1
00:00:00,000 --> 00:00:01,760
Hi, and welcome to the Neil Ashton

2
00:00:01,760 --> 00:00:04,520
podcast. In each episode, we explain

3
00:00:04,520 --> 00:00:06,080
some of the fascinating ways that

4
00:00:06,080 --> 00:00:08,200
science and engineering are changing the

5
00:00:08,200 --> 00:00:09,680
world around us.

6
00:00:09,680 --> 00:00:11,720
We talk to leading engineers from elite

7
00:00:11,720 --> 00:00:14,800
level sports like cycling and Formula 1

8
00:00:14,800 --> 00:00:16,880
to some of the world's top academics to

9
00:00:16,880 --> 00:00:19,120
understand how fluid dynamics, machine

10
00:00:19,120 --> 00:00:21,880
learning, supercomputing are bringing in

11
00:00:21,880 --> 00:00:23,840
a new era of discovery.

12
00:00:23,840 --> 00:00:26,080
We also hear some of their life stories,

13
00:00:26,080 --> 00:00:27,600
their career advice,

14
00:00:27,600 --> 00:00:29,080
and lessons they've learned on the way

15
00:00:29,080 --> 00:00:32,000
that I hope will be helpful to you, too.

16
00:00:32,000 --> 00:00:36,400
So, sit back and enjoy this episode.

17
00:00:39,760 --> 00:00:42,560
Welcome back to the Neil Ashton podcast.

18
00:00:42,560 --> 00:00:45,200
This episode is with Professor Juan

19
00:00:45,200 --> 00:00:48,800
Alonso, somebody who has

20
00:00:48,800 --> 00:00:51,640
really been a pioneer in the area of

21
00:00:51,640 --> 00:00:53,760
aeronautics and computational fluid

22
00:00:53,760 --> 00:00:56,320
dynamics.

23
00:00:56,320 --> 00:00:58,560
I thought this was a good person to

24
00:00:58,560 --> 00:01:01,840
speak to because, as I've discussed in a

25
00:01:01,840 --> 00:01:03,680
couple of previous episodes, I'm always

26
00:01:03,680 --> 00:01:06,360
fascinated by the

27
00:01:06,360 --> 00:01:08,520
by the link between academia and

28
00:01:08,520 --> 00:01:11,280
industry and the roles that that both

29
00:01:11,280 --> 00:01:14,200
play and the sort of merits of both and

30
00:01:14,200 --> 00:01:15,680
and Juan is someone who really

31
00:01:15,680 --> 00:01:17,520
epitomizes both. He's a professor at

32
00:01:17,520 --> 00:01:20,400
Stanford University, but um recently

33
00:01:20,400 --> 00:01:23,880
also founded a innovative new startup

34
00:01:23,880 --> 00:01:26,080
called Luminary Cloud.

35
00:01:26,080 --> 00:01:29,960
And we talk in the episode today

36
00:01:29,960 --> 00:01:32,880
firstly about his his career, how he got

37
00:01:32,880 --> 00:01:35,720
into engineering. I I'm I always love to

38
00:01:35,720 --> 00:01:37,840
know what motivated people, what drove

39
00:01:37,840 --> 00:01:40,240
them. I think it's so interesting to

40
00:01:40,240 --> 00:01:42,680
hear different experiences. So, we talk

41
00:01:42,680 --> 00:01:44,440
a little bit about how, you know, his

42
00:01:44,440 --> 00:01:46,440
early days, about becoming a professor,

43
00:01:46,440 --> 00:01:47,920
about, you know, leading a research

44
00:01:47,920 --> 00:01:49,040
group.

45
00:01:49,040 --> 00:01:51,760
Some of his early observations,

46
00:01:51,760 --> 00:01:52,560
um

47
00:01:52,560 --> 00:01:53,880
you know, he mentioned something like,

48
00:01:53,880 --> 00:01:56,360
"Oh, at the beginning I was didn't

49
00:01:56,360 --> 00:01:59,000
really I hadn't matured my thinking. I

50
00:01:59,000 --> 00:02:00,640
was still doing a little bit of what my

51
00:02:00,640 --> 00:02:02,440
supervisor had done. And I think that's

52
00:02:02,440 --> 00:02:03,920
true if you're an assistant professor

53
00:02:03,920 --> 00:02:05,960
straight out of your PhD, maybe you

54
00:02:05,960 --> 00:02:07,840
haven't fully, you know, got your own

55
00:02:07,840 --> 00:02:09,840
vision for everything.

56
00:02:09,840 --> 00:02:11,760
But even though he said that, if you

57
00:02:11,760 --> 00:02:14,200
look, he he was one of the first people

58
00:02:14,200 --> 00:02:15,360
to be doing a lot around high

59
00:02:15,360 --> 00:02:17,520
performance computing, looking at

60
00:02:17,520 --> 00:02:20,120
merging methods. Obviously, one of his

61
00:02:20,120 --> 00:02:22,640
big focus was on

62
00:02:22,640 --> 00:02:24,280
aircraft design.

63
00:02:24,280 --> 00:02:25,800
We didn't talk about too much. He he

64
00:02:25,800 --> 00:02:27,440
also spent some time in

65
00:02:27,440 --> 00:02:30,200
NASA headquarters. Although, as in any

66
00:02:30,200 --> 00:02:31,640
of these episodes, I feel we could have

67
00:02:31,640 --> 00:02:33,480
gone on for hours more cuz these people

68
00:02:33,480 --> 00:02:35,200
have done so much

69
00:02:35,200 --> 00:02:36,959
in their careers to date that it's hard

70
00:02:36,959 --> 00:02:38,480
to cover everything.

71
00:02:38,480 --> 00:02:40,120
But we we really get into some core

72
00:02:40,120 --> 00:02:44,200
topics around where where CFD going?

73
00:02:44,200 --> 00:02:46,320
What's the future of CFD? And one of the

74
00:02:46,320 --> 00:02:50,000
things we talk around is the CFD Vision 2030

75
00:02:50,000 --> 00:02:51,280
report.

76
00:02:51,280 --> 00:02:52,880
This was a report that Juan was one of

77
00:02:52,880 --> 00:02:55,080
the authors on, which was commissioned

78
00:02:55,080 --> 00:02:57,640
by NASA to set out

79
00:02:57,640 --> 00:03:00,760
by 2030, where do we see CFD going? And

80
00:03:00,760 --> 00:03:02,560
essentially, it talked about some grand

81
00:03:02,560 --> 00:03:05,200
challenges that need to be completed.

82
00:03:05,200 --> 00:03:06,720
So, we talked to we talked about his

83
00:03:06,720 --> 00:03:08,800
opinion on that, whether

84
00:03:08,800 --> 00:03:11,160
we already have completed some of them.

85
00:03:11,160 --> 00:03:12,840
And and so, I asked him, what would be

86
00:03:12,840 --> 00:03:15,440
the 2050 vision?

87
00:03:15,440 --> 00:03:17,720
We talked about machine learning, his

88
00:03:17,720 --> 00:03:19,720
opinion on

89
00:03:19,720 --> 00:03:20,880
you know, the hype around machine

90
00:03:20,880 --> 00:03:22,239
learning and where it's suitable and

91
00:03:22,239 --> 00:03:25,040
where it's not suitable.

92
00:03:25,040 --> 00:03:28,239
Uh and we talk about the interesting

93
00:03:28,239 --> 00:03:30,920
debate around when is open-source the

94
00:03:30,920 --> 00:03:33,000
right what can open-source codes

95
00:03:33,000 --> 00:03:34,440
achieve?

96
00:03:34,440 --> 00:03:35,880
And where

97
00:03:35,880 --> 00:03:37,560
do you need to have commercial codes?

98
00:03:37,560 --> 00:03:39,480
And I think he's uniquely placed to

99
00:03:39,480 --> 00:03:42,720
answer that question because of his work

100
00:03:42,720 --> 00:03:45,640
to help found SU2, one of the you know,

101
00:03:45,640 --> 00:03:47,720
largest open-source code that is used

102
00:03:47,720 --> 00:03:49,280
for CFD in particularly in the

103
00:03:49,280 --> 00:03:50,400
aerospace.

104
00:03:50,400 --> 00:03:53,280
And then now with Luminary Cloud, which is

105
00:03:53,280 --> 00:03:55,880
most definitely one of the disruptors in

106
00:03:55,880 --> 00:03:58,520
the CFD market, their philosophy of, you

107
00:03:58,520 --> 00:04:01,160
know, cloud first, their automated,

108
00:04:01,160 --> 00:04:04,800
super fast GPU solver is something that

109
00:04:04,800 --> 00:04:08,040
is, as we discussed, long been spoken

110
00:04:08,040 --> 00:04:09,840
about in academia,

111
00:04:09,840 --> 00:04:11,560
but only recently have their methods

112
00:04:11,560 --> 00:04:14,200
matured to be suitable for for

113
00:04:14,200 --> 00:04:17,000
commercial um code. And so he talks

114
00:04:17,000 --> 00:04:20,720
about that the motivation behind behind

115
00:04:20,720 --> 00:04:22,240
Luminary Cloud, what what made him think

116
00:04:22,240 --> 00:04:24,320
about that, the rise of cloud computing

117
00:04:24,320 --> 00:04:27,840
and HPC, the commoditization of compute.

118
00:04:27,840 --> 00:04:30,040
So we go through we go through all of it

119
00:04:30,040 --> 00:04:33,040
and I I found him a fantastic person to

120
00:04:33,040 --> 00:04:35,440
speak to. I always enjoy um seeing I

121
00:04:35,440 --> 00:04:37,520
don't see him that often because he's in

122
00:04:37,520 --> 00:04:39,560
Stanford, I'm in the UK, but whenever I

123
00:04:39,560 --> 00:04:41,720
do see him I I really

124
00:04:41,720 --> 00:04:44,040
value his judgments and his ideas and I

125
00:04:44,040 --> 00:04:46,840
hope that you also take something out of

126
00:04:46,840 --> 00:04:49,600
of of this conversation. So So I really

127
00:04:49,600 --> 00:04:52,640
hope you enjoy this episode uh too and

128
00:04:52,640 --> 00:04:54,200
if you do, make sure you, you know,

129
00:04:54,200 --> 00:04:55,880
subscribe and follow to to wherever

130
00:04:55,880 --> 00:04:57,360
you're listening to this podcast or or

131
00:04:57,360 --> 00:04:59,360
looking at it on YouTube or Spotify and

132
00:04:59,360 --> 00:05:00,360
Apple.

133
00:05:00,360 --> 00:05:02,160
Uh but yeah, sit back and enjoy this

134
00:05:02,160 --> 00:05:04,680
conversation with Professor Juan Alonso.

135
00:05:04,680 --> 00:05:06,800
First of all, thank you for doing this.

136
00:05:06,800 --> 00:05:08,680
I um

137
00:05:08,680 --> 00:05:11,000
I know it's hard to find time to speak

138
00:05:11,000 --> 00:05:12,760
to people. I'm sure with everything

139
00:05:12,760 --> 00:05:14,320
you've got now. I can only imagine how

140
00:05:14,320 --> 00:05:16,160
you're sort of juggling

141
00:05:16,160 --> 00:05:18,480
um things cuz

142
00:05:18,480 --> 00:05:21,560
I would argue that as a

143
00:05:21,560 --> 00:05:23,440
professor at Stanford,

144
00:05:23,440 --> 00:05:27,560
now a, you know, a founder of a

145
00:05:27,560 --> 00:05:30,360
pretty innovative, forward-thinking,

146
00:05:30,360 --> 00:05:32,280
potentially cutting-edge,

147
00:05:32,280 --> 00:05:33,840
groundbreaking, you know, software

148
00:05:33,840 --> 00:05:35,960
company,

149
00:05:35,960 --> 00:05:37,560
most people would say that's a pretty

150
00:05:37,560 --> 00:05:41,680
good life. How do I get to that?

151
00:05:41,720 --> 00:05:42,840
So

152
00:05:42,840 --> 00:05:46,880
did you always want to go into

153
00:05:46,880 --> 00:05:49,680
engineering and sort of aerospace

154
00:05:49,680 --> 00:05:51,080
engineering or is this one of these

155
00:05:51,080 --> 00:05:54,600
things you just fell into it?

156
00:05:55,000 --> 00:05:57,400
Well, I've always been interested in

157
00:05:57,400 --> 00:06:01,440
space and aircraft. So, my father worked

158
00:06:01,440 --> 00:06:04,880
for the Spanish airline company Iberia.

159
00:06:04,880 --> 00:06:05,200
Oh, yeah.

160
00:06:05,200 --> 00:06:07,160
And I I tell people until the sweet age

161
00:06:07,160 --> 00:06:09,440
of eight I wanted to be an astronaut and

162
00:06:09,440 --> 00:06:12,920
then I realized A, I wear glasses and B,

163
00:06:12,920 --> 00:06:14,760
I live in a country Spain that didn't

164
00:06:14,760 --> 00:06:16,280
have a space program.

165
00:06:16,280 --> 00:06:17,120
So,

166
00:06:17,120 --> 00:06:19,280
So, very quickly I decided astronaut was

167
00:06:19,280 --> 00:06:20,919
not the thing for me

168
00:06:20,919 --> 00:06:22,240
and

169
00:06:22,240 --> 00:06:24,320
I've always had a an interest in

170
00:06:24,320 --> 00:06:27,800
technology but also aesthetics. So, I

171
00:06:27,800 --> 00:06:30,200
thought of becoming an architect uh

172
00:06:30,200 --> 00:06:32,280
in Spain much like in the UK I think you

173
00:06:32,280 --> 00:06:34,560
have to select the major before you go

174
00:06:34,560 --> 00:06:36,160
into college.

175
00:06:36,160 --> 00:06:37,400
But I thought I didn't have enough

176
00:06:37,400 --> 00:06:39,480
artistic talent to be a good architect

177
00:06:39,480 --> 00:06:41,720
and I loved airplanes and I loved the

178
00:06:41,720 --> 00:06:43,400
shape and function of them and the

179
00:06:43,400 --> 00:06:45,760
engineering of it. Yeah. kind of fell

180
00:06:45,760 --> 00:06:47,640
into it but I had an interest before I

181
00:06:47,640 --> 00:06:49,680
got there.

182
00:06:49,680 --> 00:06:51,040
That you know that's so interesting you

183
00:06:51,040 --> 00:06:53,440
said that cuz I often joke with some of

184
00:06:53,440 --> 00:06:55,919
my friends that I wanted to you know

185
00:06:55,919 --> 00:06:59,000
being from the UK to be an astronaut and

186
00:06:59,000 --> 00:07:02,000
yeah, one of them was definitely the

187
00:07:02,000 --> 00:07:03,360
Oh, hold on.

188
00:07:03,360 --> 00:07:04,919
Well, the UK technically has a space

189
00:07:04,919 --> 00:07:06,200
program but

190
00:07:06,200 --> 00:07:07,800
I'm not sure I would I won't get into

191
00:07:07,800 --> 00:07:10,800
that sort of issue but

192
00:07:10,800 --> 00:07:14,160
Um but for me it was more the um I'm not

193
00:07:14,160 --> 00:07:16,960
very good with roller coasters. So,

194
00:07:16,960 --> 00:07:18,840
I kind of realized that actually

195
00:07:18,840 --> 00:07:20,160
you you have the most of them are

196
00:07:20,160 --> 00:07:21,320
fighter pilots, aren't they? So, I

197
00:07:21,320 --> 00:07:23,080
thought you have to kind of be used to

198
00:07:23,080 --> 00:07:23,960
um

199
00:07:23,960 --> 00:07:25,640
I love taking the G's.

200
00:07:25,640 --> 00:07:27,000
Oh, do you? Oh, there you go.

201
00:07:27,000 --> 00:07:29,720
Especially the ones that go upside down.

202
00:07:29,720 --> 00:07:31,640
Well, okay. So, you you are definitely

203
00:07:31,640 --> 00:07:33,440
closer. Maybe you should you could still

204
00:07:33,440 --> 00:07:35,880
do it. Do you think? Private tourism

205
00:07:35,880 --> 00:07:36,960
maybe?

206
00:07:36,960 --> 00:07:40,240
Maybe. Maybe. Someday. I hope so.

207
00:07:40,240 --> 00:07:41,600
Somebody offered me the chance to go

208
00:07:41,600 --> 00:07:43,440
into more, I'll take the the right

209
00:07:43,440 --> 00:07:46,080
anytime.

210
00:07:46,080 --> 00:07:48,160
Yeah, if you'd have asked me that

211
00:07:48,160 --> 00:07:51,080
probably 15 years ago, like when I was a

212
00:07:51,080 --> 00:07:53,040
teenager, I absolutely said yes. I think

213
00:07:53,040 --> 00:07:55,480
now coming from an engineering side and

214
00:07:55,480 --> 00:07:56,720
going, "Hmm, how often are those

215
00:07:56,720 --> 00:07:58,920
tested?" I'm

216
00:07:58,920 --> 00:08:01,320
I'm not sure.

217
00:08:01,320 --> 00:08:02,320
Um

218
00:08:02,320 --> 00:08:05,880
but okay, so you you got in more through

219
00:08:05,880 --> 00:08:07,840
the love of

220
00:08:07,840 --> 00:08:10,920
aircraft and space. So, is that what

221
00:08:10,920 --> 00:08:13,200
drove you to then go down from an

222
00:08:13,200 --> 00:08:14,880
engineering at university? Was that

223
00:08:14,880 --> 00:08:17,080
always what pushed you, you know,

224
00:08:17,080 --> 00:08:18,840
ignoring architecture, but you wanted to

225
00:08:18,840 --> 00:08:21,640
go down the engineering route?

226
00:08:21,640 --> 00:08:24,840
Yeah, I I I was analytically good,

227
00:08:24,840 --> 00:08:27,240
mathematically good. I I think I had a

228
00:08:27,240 --> 00:08:29,560
knack for physical intuition of things,

229
00:08:29,560 --> 00:08:32,120
physics. And I wanted to solve problems.

230
00:08:32,120 --> 00:08:36,640
So, I I think I thought at the time, and

231
00:08:36,640 --> 00:08:38,760
I I think it's panned out that an

232
00:08:38,760 --> 00:08:40,880
engineering career was a good career. I

233
00:08:40,880 --> 00:08:42,320
was kind of interested in airplanes and

234
00:08:42,320 --> 00:08:43,919
spacecraft, so I thought aerospace

235
00:08:43,919 --> 00:08:45,520
engineering would be good.

236
00:08:45,520 --> 00:08:48,000
And then my my career early on was a

237
00:08:48,000 --> 00:08:50,520
little bit funny. I I did my freshman

238
00:08:50,520 --> 00:08:52,360
year of college in Spain, in Madrid,

239
00:08:52,360 --> 00:08:53,839
where I'm from.

240
00:08:53,839 --> 00:08:55,120
And then I had the opportunity to come

241
00:08:55,120 --> 00:08:57,160
to the States as a sophomore, so a

242
00:08:57,160 --> 00:08:58,800
second-year student in the East in the

243
00:08:58,800 --> 00:09:00,160
States, and

244
00:09:00,160 --> 00:09:04,320
and I I ended up at MIT for my second

245
00:09:04,320 --> 00:09:06,200
year with a plan

246
00:09:06,200 --> 00:09:08,440
of trying it out for 1 year.

247
00:09:08,440 --> 00:09:10,680
And if I liked it, I would stay. If not,

248
00:09:10,680 --> 00:09:12,120
I would go back to Spain and finish my

249
00:09:12,120 --> 00:09:15,120
career there. And that was 1988, and I

250
00:09:15,120 --> 00:09:17,480
have been here ever since. So, so it was

251
00:09:17,480 --> 00:09:19,440
a great experience as an undergraduate.

252
00:09:19,440 --> 00:09:22,200
I got to learn a lot of things. I got to

253
00:09:22,200 --> 00:09:24,240
try a lot of things.

254
00:09:24,240 --> 00:09:28,000
Um both in building things and testing

255
00:09:28,000 --> 00:09:30,000
things, but also I started doing

256
00:09:30,000 --> 00:09:32,200
computations.

257
00:09:32,200 --> 00:09:34,040
Took a class from Mike Giles, who was at

258
00:09:34,040 --> 00:09:35,400
MIT at the time when I was an

259
00:09:35,400 --> 00:09:38,760
undergraduate. My first CFD class I

260
00:09:38,760 --> 00:09:40,280
think the year or two before he went

261
00:09:40,280 --> 00:09:42,440
back to Oxford. So this is must have

262
00:09:42,440 --> 00:09:45,360
been 1990 1991.

263
00:09:45,360 --> 00:09:46,880
Yeah. And

264
00:09:46,880 --> 00:09:49,640
I was sort of hooked on on the valley of

265
00:09:49,640 --> 00:09:52,240
computation and it's potential

266
00:09:52,240 --> 00:09:54,680
as it develops into the tools that you

267
00:09:54,680 --> 00:09:58,080
know we have today to really solve

268
00:09:58,080 --> 00:10:00,160
groundbreaking aerospace engineering

269
00:10:00,160 --> 00:10:02,160
problems and by extension engineering

270
00:10:02,160 --> 00:10:06,280
problems in other disciplines. So So

271
00:10:06,280 --> 00:10:07,960
Yes, there's always there's a problem

272
00:10:07,960 --> 00:10:09,840
solver in me and I thought the

273
00:10:09,840 --> 00:10:11,640
engineering profession was a good match

274
00:10:11,640 --> 00:10:13,200
to my skills.

275
00:10:13,200 --> 00:10:16,000
It always amazed me how small world I

276
00:10:16,000 --> 00:10:18,080
that Mike Giles I can't confess to know

277
00:10:18,080 --> 00:10:20,280
Mike that well but I would seem to be

278
00:10:20,280 --> 00:10:22,040
around people

279
00:10:22,040 --> 00:10:23,880
who know I was actually just having

280
00:10:23,880 --> 00:10:25,960
coffee this morning with um

281
00:10:25,960 --> 00:10:28,720
with somebody in Oxford and he he he was

282
00:10:28,720 --> 00:10:30,480
like working with him during the PhD

283
00:10:30,480 --> 00:10:32,560
years and uh

284
00:10:32,560 --> 00:10:33,960
Yeah, it's kind of funny how it works

285
00:10:33,960 --> 00:10:35,720
isn't it that you everyone seems to like

286
00:10:35,720 --> 00:10:38,400
have some connection

287
00:10:38,400 --> 00:10:38,960
to these people.

288
00:10:38,960 --> 00:10:41,200
It's a small world. He was a remarkable

289
00:10:41,200 --> 00:10:45,760
teacher and you know I I credit him for

290
00:10:45,760 --> 00:10:48,040
putting me on to CFD as as something

291
00:10:48,040 --> 00:10:49,480
that I would do as a professional later

292
00:10:49,480 --> 00:10:50,920
on. So

293
00:10:50,920 --> 00:10:53,600
But then did you So that was at MIT.

294
00:10:53,600 --> 00:10:55,080
What cuz

295
00:10:55,080 --> 00:10:56,160
I suppose one of the things I'm always

296
00:10:56,160 --> 00:10:57,920
interested in is

297
00:10:57,920 --> 00:10:59,839
as you finish your undergrad you

298
00:10:59,839 --> 00:11:01,839
obviously have an option to go into

299
00:11:01,839 --> 00:11:03,200
industry

300
00:11:03,200 --> 00:11:06,680
or academia. Was that ever a decision

301
00:11:06,680 --> 00:11:08,880
which way you were going to go?

302
00:11:08,880 --> 00:11:11,080
Oh absolutely. So you know everybody

303
00:11:11,080 --> 00:11:12,960
thinks when they see somebody who's had

304
00:11:12,960 --> 00:11:15,320
a 30 year 40 year career

305
00:11:15,320 --> 00:11:18,320
that they planned it all out. But no I I

306
00:11:18,320 --> 00:11:20,520
I was set on going back to Spain when I

307
00:11:20,520 --> 00:11:22,720
finished my studies in the US.

308
00:11:22,720 --> 00:11:24,880
And at the time when I started college

309
00:11:24,880 --> 00:11:27,520
Spain's engineering programs were six

310
00:11:27,520 --> 00:11:29,800
years six year programs. This was before

311
00:11:29,800 --> 00:11:30,680
the

312
00:11:30,680 --> 00:11:32,000
like the I think it's called the Bologna

313
00:11:32,000 --> 00:11:34,120
treaty or something where they the

314
00:11:34,120 --> 00:11:36,680
European Union and homogenized all of

315
00:11:36,680 --> 00:11:38,600
the engineering and educational

316
00:11:38,600 --> 00:11:40,920
programs. So So I knew I had to do at

317
00:11:40,920 --> 00:11:42,680
least a masters in order to be able to

318
00:11:42,680 --> 00:11:44,760
go back to Spain and just work as a

319
00:11:44,760 --> 00:11:47,280
practicing aerospace engineer.

320
00:11:47,280 --> 00:11:48,920
Yeah. So I went to grad school with that

321
00:11:48,920 --> 00:11:50,240
idea.

322
00:11:50,240 --> 00:11:51,400
Um

323
00:11:51,400 --> 00:11:53,200
I'd done a lot of engineering things,

324
00:11:53,200 --> 00:11:55,320
but I'd done little research, a little

325
00:11:55,320 --> 00:11:57,600
bit during the summers. I did not know

326
00:11:57,600 --> 00:11:58,960
that I wanted to be a researcher. In

327
00:11:58,960 --> 00:12:01,360
fact, my idea going into grad school is

328
00:12:01,360 --> 00:12:03,880
I was going to get a masters and

329
00:12:03,880 --> 00:12:05,760
potentially what I wanted to do was to

330
00:12:05,760 --> 00:12:08,200
go design airplanes in industry.

331
00:12:08,200 --> 00:12:09,240
Mhm.

332
00:12:09,240 --> 00:12:12,280
But the masters exposed me to research.

333
00:12:12,280 --> 00:12:13,760
You know, I worked with Antony Jameson at

334
00:12:13,760 --> 00:12:15,120
Princeton

335
00:12:15,120 --> 00:12:17,360
who also, you know, like Mike Giles was

336
00:12:17,360 --> 00:12:18,960
an inspiring

337
00:12:18,960 --> 00:12:22,000
Yeah. pioneer of CFD and numerical

338
00:12:22,000 --> 00:12:23,640
analysis. And

339
00:12:23,640 --> 00:12:25,960
I stayed on for the PhD and I was still

340
00:12:25,960 --> 00:12:27,360
thinking I was going to go back to

341
00:12:27,360 --> 00:12:29,240
industry and actually design airplanes.

342
00:12:29,240 --> 00:12:31,360
And And then the opportunity to come to

343
00:12:31,360 --> 00:12:33,240
Stanford as a faculty member came up and

344
00:12:33,240 --> 00:12:34,720
I thought, well, I can do that for a

345
00:12:34,720 --> 00:12:36,360
couple years. If I

346
00:12:36,360 --> 00:12:37,800
If I don't like it, I can still go and

347
00:12:37,800 --> 00:12:40,160
design airplanes. If not, it'll be very

348
00:12:40,160 --> 00:12:42,120
difficult to go the other way around. So

349
00:12:42,120 --> 00:12:43,960
So, you know, in in your career, like I

350
00:12:43,960 --> 00:12:47,160
imagine in yours, you take these make

351
00:12:47,160 --> 00:12:49,200
these decisions here and there and they

352
00:12:49,200 --> 00:12:51,160
they lead you to where you are now, but

353
00:12:51,160 --> 00:12:54,200
it was never calculated or planned out.

354
00:12:54,200 --> 00:12:56,000
Was there any of that

355
00:12:56,000 --> 00:12:58,640
of the um

356
00:12:58,640 --> 00:13:01,120
Cuz I I actually found after my PhD

357
00:13:01,120 --> 00:13:02,560
almost

358
00:13:02,560 --> 00:13:04,400
the opposite.

359
00:13:04,400 --> 00:13:06,600
I was kind of sick of academia in a way,

360
00:13:06,600 --> 00:13:08,960
you know, like it it felt like I wanted

361
00:13:08,960 --> 00:13:10,440
a change.

362
00:13:10,440 --> 00:13:12,560
Did you ever

363
00:13:12,560 --> 00:13:14,520
get down there? Or was there something

364
00:13:14,520 --> 00:13:16,400
also about moving to Stanford? Was there

365
00:13:16,400 --> 00:13:19,200
Was there or the location, living in Was

366
00:13:19,200 --> 00:13:21,000
there anything else that sort of

367
00:13:21,000 --> 00:13:22,680
did it for you?

368
00:13:22,680 --> 00:13:24,839
For me it's not as much as location. I

369
00:13:24,839 --> 00:13:26,040
It's beautiful weather here, I would

370
00:13:26,040 --> 00:13:28,040
tell you. And And it's one of the best

371
00:13:28,040 --> 00:13:29,040
weather

372
00:13:29,040 --> 00:13:31,160
sort of patterns in the world.

373
00:13:31,160 --> 00:13:32,640
Um

374
00:13:32,640 --> 00:13:35,600
to me, the thing that made me stay in

375
00:13:35,600 --> 00:13:37,920
academia, even though I have a bent

376
00:13:37,920 --> 00:13:39,720
towards solving real problems, which I

377
00:13:39,720 --> 00:13:41,680
mentioned before,

378
00:13:41,680 --> 00:13:43,640
is sort of the

379
00:13:43,640 --> 00:13:45,640
the being at the edge of the precipice

380
00:13:45,640 --> 00:13:47,360
type feeling in research when you

381
00:13:47,360 --> 00:13:48,840
actually don't know whether you're going

382
00:13:48,840 --> 00:13:50,920
to be able to find a solution for a

383
00:13:50,920 --> 00:13:53,040
problem and you're trying to do it in

384
00:13:53,040 --> 00:13:54,440
better ways than anybody's done it

385
00:13:54,440 --> 00:13:57,720
before. That was very motivating to me.

386
00:13:57,720 --> 00:13:58,320
I Yeah.

387
00:13:58,320 --> 00:14:00,840
Yeah. I think the ability to also work

388
00:14:00,840 --> 00:14:02,280
on the problems that I thought were

389
00:14:02,280 --> 00:14:04,440
important throughout my career,

390
00:14:04,440 --> 00:14:06,360
the ability to work with amazing young

391
00:14:06,360 --> 00:14:08,680
people who are far more talented than I

392
00:14:08,680 --> 00:14:10,760
am and then with whom, you know,

393
00:14:10,760 --> 00:14:13,840
together you end up doing some or or

394
00:14:13,840 --> 00:14:16,320
creating some key contributions, you

395
00:14:16,320 --> 00:14:18,160
know, to the field, that was very

396
00:14:18,160 --> 00:14:20,000
attractive to me.

397
00:14:20,000 --> 00:14:22,040
Um Mhm. I knew in an industry I could

398
00:14:22,040 --> 00:14:24,320
contribute in different ways, but I

399
00:14:24,320 --> 00:14:26,160
think my personality was better suited

400
00:14:26,160 --> 00:14:28,680
for academics and and and and to sort of

401
00:14:28,680 --> 00:14:30,200
go a little bit more into the unknown

402
00:14:30,200 --> 00:14:32,800
rather than into things that are or tend

403
00:14:32,800 --> 00:14:34,800
to be a little bit more incremental.

404
00:14:34,800 --> 00:14:35,320
Yeah.

405
00:14:35,320 --> 00:14:37,920
But So You know, if I may, in in my

406
00:14:37,920 --> 00:14:39,640
career, I've always tried to get my

407
00:14:39,640 --> 00:14:41,800
students to do something new in research

408
00:14:41,800 --> 00:14:44,000
but applied to something that is quasi

409
00:14:44,000 --> 00:14:45,880
realistic without doing the job of

410
00:14:45,880 --> 00:14:48,920
industry, of course, right? Uh Yeah.

411
00:14:48,920 --> 00:14:50,560
So it there there's been

412
00:14:50,560 --> 00:14:53,040
a constant tug-of-war between doing

413
00:14:53,040 --> 00:14:56,000
advanced new ivory tower type things and

414
00:14:56,000 --> 00:14:57,480
sort of having them be applied to

415
00:14:57,480 --> 00:14:59,520
something else and I I found that

416
00:14:59,520 --> 00:15:01,760
Stanford allowed me the opportunity to

417
00:15:01,760 --> 00:15:03,880
thrive in that sort of

418
00:15:03,880 --> 00:15:06,360
tension between the two worlds and and I

419
00:15:06,360 --> 00:15:08,920
I've enjoyed it ever since. I couldn't

420
00:15:08,920 --> 00:15:09,600
I think of myself

421
00:15:09,600 --> 00:15:12,520
Yeah, I must admit I I kind of I look up

422
00:15:12,520 --> 00:15:15,440
to you as as somebody who has done that

423
00:15:15,440 --> 00:15:18,320
very well. I think as someone Thank you.

424
00:15:18,320 --> 00:15:19,920
myself who

425
00:15:19,920 --> 00:15:21,320
feels a little bit caught between the

426
00:15:21,320 --> 00:15:23,920
two worlds that I I think what you're

427
00:15:23,920 --> 00:15:26,160
doing is actually very important because

428
00:15:26,160 --> 00:15:28,080
on one hand

429
00:15:28,080 --> 00:15:30,800
there is real value in fundamental

430
00:15:30,800 --> 00:15:33,400
science and fundamental, you know,

431
00:15:33,400 --> 00:15:36,560
theorems and and all the rest, but

432
00:15:36,560 --> 00:15:39,080
sometimes I see the sort of cynical side

433
00:15:39,080 --> 00:15:41,040
of it just, you know, for publication's

434
00:15:41,040 --> 00:15:43,200
sakes and I sometimes find myself going,

435
00:15:43,200 --> 00:15:45,000
"Yeah, but why?

436
00:15:45,000 --> 00:15:46,600
How is that going to make a difference?"

437
00:15:46,600 --> 00:15:49,560
sort of thing and Right. But if you go

438
00:15:49,560 --> 00:15:52,600
too applied, it's almost there's not the

439
00:15:52,600 --> 00:15:56,480
academic rigor and it feels like

440
00:15:56,480 --> 00:15:57,400
what you're doing different than

441
00:15:57,400 --> 00:15:59,120
industry is doing. So, it's it's a

442
00:15:59,120 --> 00:16:00,720
difficult spot I feel you're in the

443
00:16:00,720 --> 00:16:02,280
middle, isn't it?

444
00:16:02,280 --> 00:16:06,440
Yeah, so I I do think that

445
00:16:06,440 --> 00:16:08,160
you know, academia comes up with

446
00:16:08,160 --> 00:16:10,000
fundamental new ideas and sometimes they

447
00:16:10,000 --> 00:16:12,080
pan out, sometimes they don't. But any

448
00:16:12,080 --> 00:16:14,960
advanced society that already covers

449
00:16:14,960 --> 00:16:16,760
their needs for like food and shelter,

450
00:16:16,760 --> 00:16:19,720
etc., etc., goes next to trying to deal

451
00:16:19,720 --> 00:16:21,960
with knowledge and that's the creation

452
00:16:21,960 --> 00:16:24,480
and the propagation and I would say the

453
00:16:24,480 --> 00:16:26,240
perpetuation of the knowledge that has

454
00:16:26,240 --> 00:16:28,600
been acquired. So, so at the very least

455
00:16:28,600 --> 00:16:31,320
academia has value in creating the

456
00:16:31,320 --> 00:16:34,000
knowledge and perpetuating the knowledge

457
00:16:34,000 --> 00:16:36,160
by training people. I think there's much

458
00:16:36,160 --> 00:16:38,280
more value created in academia than just

459
00:16:38,280 --> 00:16:38,560
that.

460
00:16:38,560 --> 00:16:39,280
Yeah.

461
00:16:39,280 --> 00:16:41,600
But but for me in particular, it's all

462
00:16:41,600 --> 00:16:44,040
my academic ideas over 20, 30 years,

463
00:16:44,040 --> 00:16:46,160
perhaps we'll speak about this today.

464
00:16:46,160 --> 00:16:47,080
Mhm. Uh

465
00:16:47,080 --> 00:16:49,080
you know, have led to what I'm doing

466
00:16:49,080 --> 00:16:50,960
today in in sort of the entrepreneurial

467
00:16:50,960 --> 00:16:52,880
world, right? I would not have thought

468
00:16:52,880 --> 00:16:54,800
that that would happen, but without the

469
00:16:54,800 --> 00:16:57,440
academic experiences, the interactions

470
00:16:57,440 --> 00:16:59,160
with students and amazing colleagues

471
00:16:59,160 --> 00:17:01,000
that, you know, started creating

472
00:17:01,000 --> 00:17:02,960
languages so you could program in a GPU,

473
00:17:02,960 --> 00:17:06,000
for example, 20 plus years ago, you

474
00:17:06,000 --> 00:17:08,400
know, the the the company we created now

475
00:17:08,400 --> 00:17:10,920
would not have happened. So, so I I do

476
00:17:10,920 --> 00:17:12,560
think you have to navigate that balance

477
00:17:12,560 --> 00:17:13,880
and I think there are academics of all

478
00:17:13,880 --> 00:17:18,640
kinds. Um but I I think societies that

479
00:17:18,640 --> 00:17:21,880
don't have the ability to do academic

480
00:17:21,880 --> 00:17:23,160
thinking,

481
00:17:23,160 --> 00:17:26,079
innovation, creation of new ideas, and

482
00:17:26,079 --> 00:17:27,600
knowledge for the sake of knowledge,

483
00:17:27,600 --> 00:17:30,320
they end up dying. Yeah. Yeah. So, what

484
00:17:30,320 --> 00:17:31,760
was your

485
00:17:31,760 --> 00:17:33,720
you know, as I'm interested to see how

486
00:17:33,720 --> 00:17:35,800
you kind of how you've got to where

487
00:17:35,800 --> 00:17:38,440
you've got now and some of the

488
00:17:38,440 --> 00:17:40,040
route, I guess.

489
00:17:40,040 --> 00:17:41,800
So, when you were doing you were

490
00:17:41,800 --> 00:17:43,840
starting at Stanford,

491
00:17:43,840 --> 00:17:46,880
what was your core

492
00:17:46,880 --> 00:17:48,560
sort of um

493
00:17:48,560 --> 00:17:50,440
aims or sort of

494
00:17:50,440 --> 00:17:51,960
30-year goal, you know, what were what

495
00:17:51,960 --> 00:17:53,840
was your group trying to do? Cuz I feel

496
00:17:53,840 --> 00:17:55,040
like looking at some of your early

497
00:17:55,040 --> 00:17:56,720
papers, there were still things around

498
00:17:56,720 --> 00:17:59,200
high-performance computing

499
00:17:59,200 --> 00:18:01,840
very early on, arguably, in the sort of

500
00:18:01,840 --> 00:18:03,520
Sure. high-performance computing trend

501
00:18:03,520 --> 00:18:06,160
that we're in now.

502
00:18:06,160 --> 00:18:07,920
Well, so you assume there was a 30-year

503
00:18:07,920 --> 00:18:10,360
plan.

504
00:18:12,520 --> 00:18:14,840
I at the time I was very young still.

505
00:18:14,840 --> 00:18:16,520
So, I I came to Stanford straight out of

506
00:18:16,520 --> 00:18:19,480
my PhD and I had not had the opportunity

507
00:18:19,480 --> 00:18:21,840
to establish my own independence of

508
00:18:21,840 --> 00:18:23,760
thought in terms of research program.

509
00:18:23,760 --> 00:18:26,840
So, I I've been heavily influenced by my

510
00:18:26,840 --> 00:18:29,040
my PhD advisor.

511
00:18:29,040 --> 00:18:31,360
But, I also love the notion of of

512
00:18:31,360 --> 00:18:33,880
modeling the real world

513
00:18:33,880 --> 00:18:36,720
uh with mathematics and physics

514
00:18:36,720 --> 00:18:39,960
and then implementing it usefully in a

515
00:18:39,960 --> 00:18:40,840
uh

516
00:18:40,840 --> 00:18:43,240
in a digital computer in order to be

517
00:18:43,240 --> 00:18:45,400
able to to have simulators that were

518
00:18:45,400 --> 00:18:47,520
more effective than than the experiments

519
00:18:47,520 --> 00:18:49,440
and that was very early on.

520
00:18:49,440 --> 00:18:50,640
Um

521
00:18:50,640 --> 00:18:54,080
I was also smitten by the fact that I

522
00:18:54,080 --> 00:18:56,440
could see very early on that if you have

523
00:18:56,440 --> 00:18:59,000
the analysis capability, you could do

524
00:18:59,000 --> 00:19:01,000
optimization on top of the analysis to

525
00:19:01,000 --> 00:19:03,600
have tremendous potential. So, my early

526
00:19:03,600 --> 00:19:05,640
career was very much focused on getting

527
00:19:05,640 --> 00:19:07,680
the analysis in place and that was in

528
00:19:07,680 --> 00:19:09,200
numerical methods, high-performance

529
00:19:09,200 --> 00:19:12,320
computing, flow physics, etc. etc. But,

530
00:19:12,320 --> 00:19:14,400
also doing optimization on top of that.

531
00:19:14,400 --> 00:19:16,320
So, you know, I inherited

532
00:19:16,320 --> 00:19:18,840
thoughts of adjoint methods and you

533
00:19:18,840 --> 00:19:20,840
know, optimization whether single

534
00:19:20,840 --> 00:19:23,600
objective, multi-objective, etc. etc.

535
00:19:23,600 --> 00:19:25,760
And that that meant that the research

536
00:19:25,760 --> 00:19:27,760
had to grow to encompass a pretty broad

537
00:19:27,760 --> 00:19:29,840
ecosystem of various different elements.

538
00:19:29,840 --> 00:19:31,960
And I pursued some of those elements

539
00:19:31,960 --> 00:19:33,440
throughout my career. You know, some

540
00:19:33,440 --> 00:19:34,800
more successfully than others, I would

541
00:19:34,800 --> 00:19:36,360
say. But uh

542
00:19:36,360 --> 00:19:39,240
but but the plan was always to design. I

543
00:19:39,240 --> 00:19:41,080
mentioned I thought of being an aircraft

544
00:19:41,080 --> 00:19:42,400
designer.

545
00:19:42,400 --> 00:19:44,680
For for me um

546
00:19:44,680 --> 00:19:46,720
computational methods have always been a

547
00:19:46,720 --> 00:19:48,360
means to an end.

548
00:19:48,360 --> 00:19:49,880
And they still are.

549
00:19:49,880 --> 00:19:52,880
It's just that sometimes you have to

550
00:19:52,880 --> 00:19:56,200
invent, generate, or improve the tools

551
00:19:56,200 --> 00:19:57,880
in order for them to be a good means to

552
00:19:57,880 --> 00:20:00,360
an end. And and as you know, since

553
00:20:00,360 --> 00:20:01,720
you've been around for quite a bit of

554
00:20:01,720 --> 00:20:05,000
time as well, um these tools were not

555
00:20:05,000 --> 00:20:07,200
ready to be used in what I call the

556
00:20:07,200 --> 00:20:08,640
outer loops, right? You could do a

557
00:20:08,640 --> 00:20:10,440
single simulation. You could spend weeks

558
00:20:10,440 --> 00:20:12,320
and months trying to get that single

559
00:20:12,320 --> 00:20:14,480
simulation, but automatically running a

560
00:20:14,480 --> 00:20:15,960
hundred, a thousand, ten thousand

561
00:20:15,960 --> 00:20:18,920
simulations to do whatever outer loop,

562
00:20:18,920 --> 00:20:20,280
let's say optimization, but there are

563
00:20:20,280 --> 00:20:23,000
many others. That was not anything that

564
00:20:23,000 --> 00:20:24,680
could be done at the time. So, I mean,

565
00:20:24,680 --> 00:20:27,600
we were trying supersonic airplanes with

566
00:20:27,600 --> 00:20:30,240
uh Euler methods. You know, we were

567
00:20:30,240 --> 00:20:32,640
trying genetic algorithms. We were

568
00:20:32,640 --> 00:20:34,400
trying all kinds of interesting things

569
00:20:34,400 --> 00:20:35,680
to see

570
00:20:35,680 --> 00:20:38,920
what would work or what methods would

571
00:20:38,920 --> 00:20:41,640
work in which applications. And that was

572
00:20:41,640 --> 00:20:44,120
my early career. And the idea was I

573
00:20:44,120 --> 00:20:46,080
wanted to get everything ready so I

574
00:20:46,080 --> 00:20:48,840
could do optimization.

575
00:20:48,840 --> 00:20:51,320
Based design as I can understand and and

576
00:20:51,320 --> 00:20:53,960
simulation-based optimization.

577
00:20:53,960 --> 00:20:55,600
So, that that that was what motivated me

578
00:20:55,600 --> 00:20:57,040
in the early days.

579
00:20:57,040 --> 00:20:58,520
And there was a lot of America houses

580
00:20:58,520 --> 00:21:00,640
and a lot of parallel computing. Yeah.

581
00:21:00,640 --> 00:21:02,800
How do you think it's changed then? So,

582
00:21:02,800 --> 00:21:05,160
if you look back when in when was it

583
00:21:05,160 --> 00:21:08,360
night well, no, early 2000s.

584
00:21:08,360 --> 00:21:09,440
Yeah.

585
00:21:09,440 --> 00:21:10,680
Um

586
00:21:10,680 --> 00:21:12,120
if you were to go to an aircraft

587
00:21:12,120 --> 00:21:13,640
manufacturer,

588
00:21:13,640 --> 00:21:17,040
look at their design process then,

589
00:21:17,040 --> 00:21:19,560
and fast forward it to now,

590
00:21:19,560 --> 00:21:21,400
what would you see as being the sort of

591
00:21:21,400 --> 00:21:22,920
big

592
00:21:22,920 --> 00:21:25,800
differences?

593
00:21:26,680 --> 00:21:27,920
Um

594
00:21:27,920 --> 00:21:29,680
I don't want this to sound critical of

595
00:21:29,680 --> 00:21:31,720
the aerospace industry cuz this safety

596
00:21:31,720 --> 00:21:34,640
critical industry that requires

597
00:21:34,640 --> 00:21:36,920
very careful attention paid to to large

598
00:21:36,920 --> 00:21:40,000
steps, large changes in design

599
00:21:40,000 --> 00:21:42,160
processes, and manufacturing process

600
00:21:42,160 --> 00:21:44,640
control systems, whatever the heck.

601
00:21:44,640 --> 00:21:48,040
Um I think the use of computation

602
00:21:48,040 --> 00:21:50,120
from when I got out of grad school, so

603
00:21:50,120 --> 00:21:53,160
in '97, until now,

604
00:21:53,160 --> 00:21:54,120
it's gotten more automated and it's

605
00:21:54,120 --> 00:21:55,400
gotten faster,

606
00:21:55,400 --> 00:21:56,640
but it's not gotten substantially

607
00:21:56,640 --> 00:21:58,520
different. There are a few companies

608
00:21:58,520 --> 00:22:01,400
around the world who try to do some of

609
00:22:01,400 --> 00:22:03,280
the outer loops, sometimes more

610
00:22:03,280 --> 00:22:05,080
successfully than others,

611
00:22:05,080 --> 00:22:08,040
seldom in production, I would say, most

612
00:22:08,040 --> 00:22:10,640
often than not in their R&D centers,

613
00:22:10,640 --> 00:22:12,520
sort of informing studies and various

614
00:22:12,520 --> 00:22:13,880
other things. So

615
00:22:13,880 --> 00:22:14,920
So

616
00:22:14,920 --> 00:22:16,800
surprisingly, we've come up with a

617
00:22:16,800 --> 00:22:18,800
number of technologies to accelerate the

618
00:22:18,800 --> 00:22:20,480
simulations,

619
00:22:20,480 --> 00:22:22,680
and we come up, I would say, with a

620
00:22:22,680 --> 00:22:24,920
number of technologies when you're doing

621
00:22:24,920 --> 00:22:26,760
the same thing over and over again to

622
00:22:26,760 --> 00:22:30,280
automate, script, link things together,

623
00:22:30,280 --> 00:22:32,040
which is one big part of engineering,

624
00:22:32,040 --> 00:22:33,600
don't take me wrong, but I I don't think

625
00:22:33,600 --> 00:22:36,120
we fundamentally talked about changing

626
00:22:36,120 --> 00:22:38,280
the way in which we either design

627
00:22:38,280 --> 00:22:41,160
systems or use the computational tools

628
00:22:41,160 --> 00:22:43,200
to design those systems. And I think in

629
00:22:43,200 --> 00:22:45,280
academia we we have more freedom to

630
00:22:45,280 --> 00:22:47,720
think about these things, and I tried to

631
00:22:47,720 --> 00:22:50,080
devote a lot of my academic experience

632
00:22:50,080 --> 00:22:52,600
to how do you reimagine the way you do

633
00:22:52,600 --> 00:22:53,840
design,

634
00:22:53,840 --> 00:22:56,400
and and more recently in in Luminary,

635
00:22:56,400 --> 00:22:58,960
the the small startup that we created

636
00:22:58,960 --> 00:23:01,880
about 4 and 1/2 years ago or so, it's

637
00:23:01,880 --> 00:23:04,400
it's it's been about saying, "Okay, we

638
00:23:04,400 --> 00:23:06,880
know what we need to do,

639
00:23:06,880 --> 00:23:08,800
but we got to put it together, right?"

640
00:23:08,800 --> 00:23:10,200
And it's it's

641
00:23:10,200 --> 00:23:11,800
it's interesting. I I don't know what

642
00:23:11,800 --> 00:23:13,680
you think about this, and I don't mean

643
00:23:13,680 --> 00:23:15,280
to interview you in your own podcast,

644
00:23:15,280 --> 00:23:16,360
but

645
00:23:16,360 --> 00:23:18,400
but it's been very clear to me from the

646
00:23:18,400 --> 00:23:21,160
beginning that what engineers need is

647
00:23:21,160 --> 00:23:23,000
that inner loop, you know, going from

648
00:23:23,000 --> 00:23:25,880
geometry to output of simulation to be

649
00:23:25,880 --> 00:23:28,800
very fast, guaranteed level of accuracy,

650
00:23:28,800 --> 00:23:30,600
whatever is requested,

651
00:23:30,600 --> 00:23:32,800
very robust, meaning it always runs

652
00:23:32,800 --> 00:23:34,760
without worrying about this mesh or that

653
00:23:34,760 --> 00:23:36,840
mesh or this parameter, and very

654
00:23:36,840 --> 00:23:39,120
scalably, so you can sometimes run five

655
00:23:39,120 --> 00:23:41,080
or 10 at a time and sometimes 100 at a

656
00:23:41,080 --> 00:23:43,800
time, right? And that's end to end.

657
00:23:43,800 --> 00:23:46,000
So so it's become very clear over the

658
00:23:46,000 --> 00:23:48,480
years that if you had that capability,

659
00:23:48,480 --> 00:23:50,040
if you could put a geometry and you get

660
00:23:50,040 --> 00:23:53,640
an answer fast, accurately, reliably,

661
00:23:53,640 --> 00:23:55,840
you know, and scalably, you'd be able to

662
00:23:55,840 --> 00:23:58,120
then put all these outer loops,

663
00:23:58,120 --> 00:23:59,600
optimization, uncertainty

664
00:23:59,600 --> 00:24:01,800
quantification, parameter studies, you

665
00:24:01,800 --> 00:24:05,080
know, AI, ML, so on and so forth. The

666
00:24:05,080 --> 00:24:07,480
essence is that inner loop. The way we

667
00:24:07,480 --> 00:24:09,720
do the outer loops is is relatively well

668
00:24:09,720 --> 00:24:12,120
known at the moment, and and we just

669
00:24:12,120 --> 00:24:13,520
need to make it work and make it work

670
00:24:13,520 --> 00:24:16,520
fast. So So I would say I've I've

671
00:24:16,520 --> 00:24:18,840
devoted a lot of my career and more

672
00:24:18,840 --> 00:24:21,560
recently, you know, Luminary, my time

673
00:24:21,560 --> 00:24:23,080
and effort to make sure that we get

674
00:24:23,080 --> 00:24:25,400
those inner loops nailed down,

675
00:24:25,400 --> 00:24:27,840
so all the outer loops become possible.

676
00:24:27,840 --> 00:24:29,000
I don't know if that makes sense to you,

677
00:24:29,000 --> 00:24:29,800
but

678
00:24:29,800 --> 00:24:31,120
No, no, no.

679
00:24:31,120 --> 00:24:33,960
I think um

680
00:24:34,280 --> 00:24:37,160
I I I think the bit that I'm interested

681
00:24:37,160 --> 00:24:40,760
in it does seem as if the CFD, I mean,

682
00:24:40,760 --> 00:24:43,080
fast forwarding a little bit to now,

683
00:24:43,080 --> 00:24:43,920
Yeah.

684
00:24:43,920 --> 00:24:48,640
the CFD world, it did feel in a way was

685
00:24:48,640 --> 00:24:51,800
almost a little bit static. And I know

686
00:24:51,800 --> 00:24:53,040
some people would say, "No, that's not

687
00:24:53,040 --> 00:24:54,960
true." But it felt

688
00:24:54,960 --> 00:24:57,200
it was a period,

689
00:24:57,200 --> 00:24:59,880
let's say before COVID,

690
00:24:59,880 --> 00:25:01,320
where if I was to look at the codes that

691
00:25:01,320 --> 00:25:03,400
were coming out and everything, it was

692
00:25:03,400 --> 00:25:05,880
mainly CPU-based,

693
00:25:05,880 --> 00:25:07,280
mainly

694
00:25:07,280 --> 00:25:09,200
still RANS-

695
00:25:09,200 --> 00:25:12,960
based, unstructured grids, and

696
00:25:12,960 --> 00:25:15,040
people were tweaking things, but if if

697
00:25:15,040 --> 00:25:16,440
you looked

698
00:25:16,440 --> 00:25:18,200
from the early 2000s until like the

699
00:25:18,200 --> 00:25:20,800
2020s, people were using models still

700
00:25:20,800 --> 00:25:23,160
published in the 1990s.

701
00:25:23,160 --> 00:25:25,720
There wasn't that much. And it well,

702
00:25:25,720 --> 00:25:27,160
maybe that were happening were in

703
00:25:27,160 --> 00:25:28,960
academia, but they weren't in like

704
00:25:28,960 --> 00:25:30,720
industry production. Where if you fast

705
00:25:30,720 --> 00:25:31,920
forward

706
00:25:31,920 --> 00:25:33,520
to now,

707
00:25:33,520 --> 00:25:35,800
I'm seeing

708
00:25:35,800 --> 00:25:38,200
companies like yours,

709
00:25:38,200 --> 00:25:41,080
and to be fair, a broader sort of like

710
00:25:41,080 --> 00:25:42,640
emergence

711
00:25:42,640 --> 00:25:45,200
of these new companies.

712
00:25:45,200 --> 00:25:46,560
What

713
00:25:46,560 --> 00:25:47,960
What do you think is driving that? Like

714
00:25:47,960 --> 00:25:49,320
what made you do it? What made it

715
00:25:49,320 --> 00:25:51,640
possible for you to do it in,

716
00:25:51,640 --> 00:25:54,520
you know, now in the past 4 years versus

717
00:25:54,520 --> 00:25:56,240
50 years ago? Like what what what's made

718
00:25:56,240 --> 00:25:59,800
the conditions right for it to happen?

719
00:25:59,800 --> 00:26:01,720
I I So, I would agree with you. You

720
00:26:01,720 --> 00:26:04,560
know, up to about 5 6 7 years ago,

721
00:26:04,560 --> 00:26:06,960
except in academia, where we were all

722
00:26:06,960 --> 00:26:08,880
trying all kinds of crazy interesting

723
00:26:08,880 --> 00:26:11,440
things, Mhm. nothing much had happened

724
00:26:11,440 --> 00:26:13,920
other than more cores,

725
00:26:13,920 --> 00:26:17,480
same code, more MPI ranks, you know,

726
00:26:17,480 --> 00:26:19,760
a few automations here and there, mesh

727
00:26:19,760 --> 00:26:21,840
generation for unstructured methods, you

728
00:26:21,840 --> 00:26:23,760
know, becoming a more mature discipline,

729
00:26:23,760 --> 00:26:26,840
although, you still in in need of lots

730
00:26:26,840 --> 00:26:27,600
of

731
00:26:27,600 --> 00:26:29,320
sort of interesting improvements, I

732
00:26:29,320 --> 00:26:31,720
would say. And and and very stagnated

733
00:26:31,720 --> 00:26:35,120
field. Um

734
00:26:35,480 --> 00:26:38,320
To me, and I if I switch now to my

735
00:26:38,320 --> 00:26:39,840
non-academic and and more

736
00:26:39,840 --> 00:26:42,880
entrepreneurial Luminary career,

737
00:26:42,880 --> 00:26:45,280
um in the 2019

738
00:26:45,280 --> 00:26:47,520
time frame, when I was talking to the

739
00:26:47,520 --> 00:26:49,520
person who'll become my co-founder here

740
00:26:49,520 --> 00:26:51,040
at Luminary,

741
00:26:51,040 --> 00:26:53,320
it become very clear that a number of

742
00:26:53,320 --> 00:26:55,920
the key technologies that were needed to

743
00:26:55,920 --> 00:26:58,480
realize the vision were now mature

744
00:26:58,480 --> 00:27:00,120
enough that you could actually have a

745
00:27:00,120 --> 00:27:02,920
chance of getting it done in a

746
00:27:02,920 --> 00:27:05,320
short amount of time. I put short in

747
00:27:05,320 --> 00:27:07,880
between quotes, right? So,

748
00:27:07,880 --> 00:27:09,760
um

749
00:27:09,760 --> 00:27:12,560
we started looking at sort of GPU

750
00:27:12,560 --> 00:27:14,800
computing at Stanford. You know, these

751
00:27:14,800 --> 00:27:16,160
are colleagues in CS that were

752
00:27:16,160 --> 00:27:18,560
participating these ASCII programs for

753
00:27:18,560 --> 00:27:21,000
full jet engine simulations.

754
00:27:21,000 --> 00:27:24,480
In in the early 2000s, this is prior to

755
00:27:24,480 --> 00:27:26,760
CUDA, etc. etc. In fact, it was a

756
00:27:26,760 --> 00:27:29,160
student of Pat Hanrahan's I met I I

757
00:27:29,160 --> 00:27:31,760
remember Ian Buck, who's now at NVIDIA,

758
00:27:31,760 --> 00:27:34,760
who had done something called BrookGPU

759
00:27:34,760 --> 00:27:36,600
where the texture memory of the existing

760
00:27:36,600 --> 00:27:38,560
GPUs of the time was used as memory for

761
00:27:38,560 --> 00:27:41,640
computations. And and Ian then got hired

762
00:27:41,640 --> 00:27:43,400
at NVIDIA and and became the main

763
00:27:43,400 --> 00:27:45,040
architect of CUDA, right? And of course,

764
00:27:45,040 --> 00:27:46,600
at the same time NVIDIA started changing

765
00:27:46,600 --> 00:27:48,560
the hardware of the GPUs in order to

766
00:27:48,560 --> 00:27:50,880
enable computing. So, so back in the

767
00:27:50,880 --> 00:27:53,160
early 2000s, we were already doing these

768
00:27:53,160 --> 00:27:55,520
things. In fact, I remember with another

769
00:27:55,520 --> 00:27:56,960
colleague from Stanford, who's also at

770
00:27:56,960 --> 00:28:00,000
NVIDIA now, Massimiliano Fatica, in

771
00:28:00,000 --> 00:28:02,600
2004, we published two papers in the

772
00:28:02,600 --> 00:28:04,640
AIAA Aerospace Sciences Meeting, one

773
00:28:04,640 --> 00:28:07,040
called Stream Flow and the other one was

774
00:28:07,040 --> 00:28:08,840
Stream FEM.

775
00:28:08,840 --> 00:28:12,080
The The GPU computing name back then was

776
00:28:12,080 --> 00:28:13,840
coined by Bill Dally, who's now a chief

777
00:28:13,840 --> 00:28:15,760
scientist at NVIDIA, at the time a

778
00:28:15,760 --> 00:28:18,200
professor in CS as well, and electrical

779
00:28:18,200 --> 00:28:19,600
engineering, I think.

780
00:28:19,600 --> 00:28:22,440
Um the trea- the term was coined stream

781
00:28:22,440 --> 00:28:24,040
computing, right? So, or stream

782
00:28:24,040 --> 00:28:25,600
supercomputing, so.

783
00:28:25,600 --> 00:28:26,560
So,

784
00:28:26,560 --> 00:28:30,040
back then, those ideas were germinating

785
00:28:30,040 --> 00:28:32,040
and now they were ready in let's say the

786
00:28:32,040 --> 00:28:34,680
2018-2019 timeframe.

787
00:28:34,680 --> 00:28:38,800
Cloud was unheard of in the early 2000s.

788
00:28:38,800 --> 00:28:40,880
You know this better than I do. But you

789
00:28:40,880 --> 00:28:42,960
may remember grid computing, right? Or

790
00:28:42,960 --> 00:28:45,240
very different supercomputing. So, the

791
00:28:45,240 --> 00:28:47,480
the idea of the cloud was germinating

792
00:28:47,480 --> 00:28:49,680
back then as well, but you know, the

793
00:28:49,680 --> 00:28:52,600
costs, the high availability, the

794
00:28:52,600 --> 00:28:55,400
security, and various other, you know,

795
00:28:55,400 --> 00:28:57,400
virtualization layers that came from

796
00:28:57,400 --> 00:28:59,600
many of those early research projects,

797
00:28:59,600 --> 00:29:02,200
you know, were also panning out. The

798
00:29:02,200 --> 00:29:04,040
infrastructure wasn't place, so you

799
00:29:04,040 --> 00:29:06,080
could actually run a calculation at data

800
00:29:06,080 --> 00:29:08,280
center in the middle of the US and sort

801
00:29:08,280 --> 00:29:10,680
of visualize it in real time from a

802
00:29:10,680 --> 00:29:13,400
browser, you know, in the Bay Area. So,

803
00:29:13,400 --> 00:29:15,720
all of these things were coming together

804
00:29:15,720 --> 00:29:18,280
and it was clear that it was possible to

805
00:29:18,280 --> 00:29:20,080
do something different.

806
00:29:20,080 --> 00:29:22,120
Um

807
00:29:22,120 --> 00:29:24,400
Yeah, it it became very clear that that

808
00:29:24,400 --> 00:29:26,960
was the solution. That eventually most

809
00:29:26,960 --> 00:29:28,360
of the high-performance computing in the

810
00:29:28,360 --> 00:29:31,480
world will go away from on-prem clusters

811
00:29:31,480 --> 00:29:33,640
and and just go into the cloud. And and

812
00:29:33,640 --> 00:29:35,440
I know you work for one such cloud

813
00:29:35,440 --> 00:29:37,720
service provider, so I I know you're

814
00:29:37,720 --> 00:29:39,240
going to agree with me, but the

815
00:29:39,240 --> 00:29:41,440
competition between the major cloud

816
00:29:41,440 --> 00:29:43,200
service providers over the next 10 years

817
00:29:43,200 --> 00:29:44,920
is going to be fierce, which means

818
00:29:44,920 --> 00:29:46,600
prices will come down, capabilities will

819
00:29:46,600 --> 00:29:48,760
go up, and additional differentiators

820
00:29:48,760 --> 00:29:50,200
will will be attempted by various

821
00:29:50,200 --> 00:29:52,480
different places. So, so there'll be a

822
00:29:52,480 --> 00:29:55,280
continuous drive for innovation. So, to

823
00:29:55,280 --> 00:29:57,200
to me it became very clear. I've been

824
00:29:57,200 --> 00:29:59,360
doing some things in the cloud earlier

825
00:29:59,360 --> 00:30:02,680
on in my research program and saw how

826
00:30:02,680 --> 00:30:03,920
good it was.

827
00:30:03,920 --> 00:30:05,560
So, it became very clear that a

828
00:30:05,560 --> 00:30:08,320
combination of GPU computing native from

829
00:30:08,320 --> 00:30:10,800
zero, you know, the beginning, and cloud

830
00:30:10,800 --> 00:30:13,240
computing native from the beginning

831
00:30:13,240 --> 00:30:15,520
was the way to have that sort of step

832
00:30:15,520 --> 00:30:18,320
function in terms of improvement of how

833
00:30:18,320 --> 00:30:19,960
computational simulations could be used

834
00:30:19,960 --> 00:30:22,840
for analysis and design, right? So, So,

835
00:30:22,840 --> 00:30:24,800
that's a long-winded answer to your

836
00:30:24,800 --> 00:30:27,480
question, but but that's precisely what

837
00:30:27,480 --> 00:30:29,560
was happening. It was very clear there

838
00:30:29,560 --> 00:30:31,880
was a confluence of things

839
00:30:31,880 --> 00:30:34,000
happening that will make this possible.

840
00:30:34,000 --> 00:30:35,680
And

841
00:30:35,680 --> 00:30:36,680
go ahead.

842
00:30:36,680 --> 00:30:38,320
No, I was going to say

843
00:30:38,320 --> 00:30:39,880
and it's not like we had a, you know,

844
00:30:39,880 --> 00:30:41,440
crystal ball and everything was

845
00:30:41,440 --> 00:30:43,880
perfectly crystal clear, you know. There

846
00:30:43,880 --> 00:30:46,040
was con- There were concerns about

847
00:30:46,040 --> 00:30:48,440
companies putting their IP in the cloud,

848
00:30:48,440 --> 00:30:50,840
you know, data security issues,

849
00:30:50,840 --> 00:30:54,200
performance issues, you know,

850
00:30:54,200 --> 00:30:56,720
the the level of differentiation that

851
00:30:56,720 --> 00:30:59,120
one could achieve versus existing

852
00:30:59,120 --> 00:31:01,600
software providers, uh and and whether

853
00:31:01,600 --> 00:31:03,760
we could be successful, and frankly, how

854
00:31:03,760 --> 00:31:05,560
long it would take to redevelop

855
00:31:05,560 --> 00:31:07,840
everything from scratch. Not just for

856
00:31:07,840 --> 00:31:11,040
solver technology, meshing, adaptation,

857
00:31:11,040 --> 00:31:13,240
visualization, AI machine learning,

858
00:31:13,240 --> 00:31:16,600
right? CAD ingestion, CAD cleanup, um

859
00:31:16,600 --> 00:31:18,320
you know, interpretations of results,

860
00:31:18,320 --> 00:31:20,520
creation of databases, scaling to large

861
00:31:20,520 --> 00:31:22,760
numbers of users, you know, scaling to

862
00:31:22,760 --> 00:31:24,600
large numbers of GPU resources in the

863
00:31:24,600 --> 00:31:27,160
cloud. Those were all unknowns that one

864
00:31:27,160 --> 00:31:28,640
needed to come up with solutions for

865
00:31:28,640 --> 00:31:31,200
that were less in the academic realm and

866
00:31:31,200 --> 00:31:34,080
more in the, I would say, creativity and

867
00:31:34,080 --> 00:31:36,680
innovation realm, right?

868
00:31:36,680 --> 00:31:39,080
How important do you think it

869
00:31:39,080 --> 00:31:41,400
it is

870
00:31:41,400 --> 00:31:43,480
that you are where you are in the world?

871
00:31:43,480 --> 00:31:44,560
Like, do you think you could have

872
00:31:44,560 --> 00:31:47,280
attracted the investment on the idea if

873
00:31:47,280 --> 00:31:49,600
you weren't in the Bay Area? Is there

874
00:31:49,600 --> 00:31:51,720
something about the Bay Area that helps,

875
00:31:51,720 --> 00:31:52,280
or

876
00:31:52,280 --> 00:31:53,960
Yeah.

877
00:31:53,960 --> 00:31:56,320
Yeah. Yeah, I

878
00:31:56,320 --> 00:31:57,320
I

879
00:31:57,320 --> 00:31:59,440
I mean,

880
00:31:59,440 --> 00:32:01,000
I don't have

881
00:32:01,000 --> 00:32:03,480
proof that the following statement is

882
00:32:03,480 --> 00:32:08,240
true or factual, but I have a strong

883
00:32:08,240 --> 00:32:10,480
inclination to say that it would have

884
00:32:10,480 --> 00:32:12,200
been almost impossible anywhere in the

885
00:32:12,200 --> 00:32:13,240
world.

886
00:32:13,240 --> 00:32:15,080
You know,

887
00:32:15,080 --> 00:32:17,480
um you know, how

888
00:32:17,480 --> 00:32:19,840
how much work goes into just doing a

889
00:32:19,840 --> 00:32:22,200
solver. So, now try to do solvers for

890
00:32:22,200 --> 00:32:23,840
compressible and incompressible. Try to

891
00:32:23,840 --> 00:32:26,440
add porous media. Try to add multi-phase

892
00:32:26,440 --> 00:32:28,720
flows, thermal solvers, aeroacoustics,

893
00:32:28,720 --> 00:32:31,160
etc., etc., eventually structures. It's

894
00:32:31,160 --> 00:32:33,160
a massive amount of effort, and it

895
00:32:33,160 --> 00:32:35,720
requires a very large number of people

896
00:32:35,720 --> 00:32:38,800
by academic standards. So, it requires

897
00:32:38,800 --> 00:32:40,640
funding because these things don't get

898
00:32:40,640 --> 00:32:43,760
done in 6 months, right? We're not I I

899
00:32:43,760 --> 00:32:45,400
remind people, we're not putting

900
00:32:45,400 --> 00:32:47,240
together an online calendaring

901
00:32:47,240 --> 00:32:48,600
application.

902
00:32:48,600 --> 00:32:50,720
You know it's going to work, right?

903
00:32:50,720 --> 00:32:51,520
You

904
00:32:51,520 --> 00:32:52,680
you're trying to do something that's

905
00:32:52,680 --> 00:32:54,920
never been done before and you want to

906
00:32:54,920 --> 00:32:56,560
make sure it's accurate, it's fast, it

907
00:32:56,560 --> 00:32:59,240
is robust, it works all the time, right?

908
00:32:59,240 --> 00:33:01,560
Etc. etc. So, so that requires a certain

909
00:33:01,560 --> 00:33:02,720
level of investment that would have been

910
00:33:02,720 --> 00:33:04,240
hard to find anywhere else but in the

911
00:33:04,240 --> 00:33:06,120
Bay Area.

912
00:33:06,120 --> 00:33:08,880
Also in Luminary, unlike in my academic

913
00:33:08,880 --> 00:33:11,280
research, there are two parts of the

914
00:33:11,280 --> 00:33:12,320
company.

915
00:33:12,320 --> 00:33:14,040
One

916
00:33:14,040 --> 00:33:15,760
are software developer computer

917
00:33:15,760 --> 00:33:17,480
scientists

918
00:33:17,480 --> 00:33:19,760
that are highly concentrated in this

919
00:33:19,760 --> 00:33:21,400
area of the world and that are essential

920
00:33:21,400 --> 00:33:23,240
to the success of the company.

921
00:33:23,240 --> 00:33:24,960
The other are people like you and me,

922
00:33:24,960 --> 00:33:26,200
sort of trained as computational

923
00:33:26,200 --> 00:33:28,280
scientists, visualization experts, AI

924
00:33:28,280 --> 00:33:30,560
machine learning, meshing, geometry,

925
00:33:30,560 --> 00:33:32,160
etc. etc.

926
00:33:32,160 --> 00:33:35,040
Um and those exist all over the world.

927
00:33:35,040 --> 00:33:36,440
And in fact, you have to attract them to

928
00:33:36,440 --> 00:33:37,600
a place

929
00:33:37,600 --> 00:33:39,560
where they want to be. And the Bay Area

930
00:33:39,560 --> 00:33:41,160
is a nice place to attract people

931
00:33:41,160 --> 00:33:43,520
despite the housing costs.

932
00:33:43,520 --> 00:33:44,680
Um

933
00:33:44,680 --> 00:33:46,120
but but yes, I think it would have been

934
00:33:46,120 --> 00:33:47,760
very difficult to do anywhere else

935
00:33:47,760 --> 00:33:49,920
because of the magnitude of the

936
00:33:49,920 --> 00:33:52,320
investment, the time that it takes and

937
00:33:52,320 --> 00:33:54,320
sustain level that it requires to

938
00:33:54,320 --> 00:33:56,400
actually get there. And I would say

939
00:33:56,400 --> 00:33:58,320
because you're going to have to attract

940
00:33:58,320 --> 00:34:00,960
talented people from across the world.

941
00:34:00,960 --> 00:34:02,760
We have a good pool of software

942
00:34:02,760 --> 00:34:04,400
developers, but everybody else has to

943
00:34:04,400 --> 00:34:06,720
come from wherever they are. So, they're

944
00:34:06,720 --> 00:34:09,159
the best in what they do and and

945
00:34:09,159 --> 00:34:11,440
attracting them here versus

946
00:34:11,440 --> 00:34:13,080
other places in the world is is a little

947
00:34:13,080 --> 00:34:15,720
bit easier, right? Yeah.

948
00:34:15,720 --> 00:34:17,399
How much

949
00:34:17,399 --> 00:34:18,639
do you

950
00:34:18,639 --> 00:34:20,120
I don't know how to phrase this.

951
00:34:20,120 --> 00:34:23,080
Do you have more respect now

952
00:34:23,080 --> 00:34:25,320
not that you were disrespecting for some

953
00:34:25,320 --> 00:34:27,520
of the large ISV companies. Like I

954
00:34:27,520 --> 00:34:29,600
always got this sense in academia that

955
00:34:29,600 --> 00:34:32,200
sometimes it was easy to

956
00:34:32,200 --> 00:34:34,399
criticize large commercial companies and

957
00:34:34,399 --> 00:34:35,800
be like, oh, you know, their codes are

958
00:34:35,800 --> 00:34:37,120
not very good. You know, what we've

959
00:34:37,120 --> 00:34:39,120
developed in our paper is better. Having

960
00:34:39,120 --> 00:34:40,760
to do what you've done, do you suddenly

961
00:34:40,760 --> 00:34:42,800
go, hmm, actually now I see it's

962
00:34:42,800 --> 00:34:45,639
actually quite hard to build up these

963
00:34:45,639 --> 00:34:47,520
big code bases and

964
00:34:47,520 --> 00:34:51,600
validate them, etc. Yeah, um

965
00:34:51,879 --> 00:34:54,840
Yes and no. So, two things, of course,

966
00:34:54,840 --> 00:34:56,200
um

967
00:34:56,200 --> 00:34:58,360
Yes, producing a product that can be

968
00:34:58,360 --> 00:34:59,920
applied to many different types of

969
00:34:59,920 --> 00:35:02,520
applications and that are willing to pay

970
00:35:02,520 --> 00:35:05,360
for because it's it's addressing pain

971
00:35:05,360 --> 00:35:07,400
that they have in the current processes

972
00:35:07,400 --> 00:35:08,720
is very hard.

973
00:35:08,720 --> 00:35:10,720
It's much harder than writing an open

974
00:35:10,720 --> 00:35:12,840
source software solver and and making

975
00:35:12,840 --> 00:35:15,000
sure that, you know, academics with a

976
00:35:15,000 --> 00:35:17,000
lot of patience and and some industry

977
00:35:17,000 --> 00:35:19,200
and and government people are able to

978
00:35:19,200 --> 00:35:21,040
use it and and they ask questions and

979
00:35:21,040 --> 00:35:22,520
it's okay they didn't pay anything for

980
00:35:22,520 --> 00:35:24,080
it. So, you know, if it doesn't work as

981
00:35:24,080 --> 00:35:26,560
advertised, it's all right. So, gain a

982
00:35:26,560 --> 00:35:28,480
tremendous amount of respect for

983
00:35:28,480 --> 00:35:31,080
companies that put good, high-quality

984
00:35:31,080 --> 00:35:32,640
products together

985
00:35:32,640 --> 00:35:33,960
because that means there's a tremendous

986
00:35:33,960 --> 00:35:36,520
amount of thinking about the features,

987
00:35:36,520 --> 00:35:38,160
how they're exposed to the users, how

988
00:35:38,160 --> 00:35:39,680
they're implemented, what type of

989
00:35:39,680 --> 00:35:41,760
regression testing you do, you know,

990
00:35:41,760 --> 00:35:43,680
what type of user testing you do, how do

991
00:35:43,680 --> 00:35:45,480
you make sure that you're doing

992
00:35:45,480 --> 00:35:47,240
something for the sake of improving

993
00:35:47,240 --> 00:35:49,360
processes. So, so that takes an

994
00:35:49,360 --> 00:35:50,880
inordinate amount of time. You know, if

995
00:35:50,880 --> 00:35:52,360
you develop something in academia and

996
00:35:52,360 --> 00:35:54,080
you think you're done, you've done about

997
00:35:54,080 --> 00:35:56,480
10% of what it is required to actually

998
00:35:56,480 --> 00:35:58,640
put together a viable product. So,

999
00:35:58,640 --> 00:36:01,360
So, in that sense, yes, I I I have a

1000
00:36:01,360 --> 00:36:03,000
tremendous amount of respect for those

1001
00:36:03,000 --> 00:36:04,360
companies that actually do this

1002
00:36:04,360 --> 00:36:05,560
properly.

1003
00:36:05,560 --> 00:36:06,000
Yeah.

1004
00:36:06,000 --> 00:36:06,880
Um

1005
00:36:06,880 --> 00:36:09,040
The converse is also true.

1006
00:36:09,040 --> 00:36:10,120
You know, some of these companies, I

1007
00:36:10,120 --> 00:36:12,040
mean, they're around for a long, long

1008
00:36:12,040 --> 00:36:14,000
time.

1009
00:36:14,000 --> 00:36:17,480
And while their solver technology has

1010
00:36:17,480 --> 00:36:19,800
improved over the years, they missed

1011
00:36:19,800 --> 00:36:22,840
opportunities over time

1012
00:36:22,840 --> 00:36:24,640
to really do much more significant

1013
00:36:24,640 --> 00:36:26,360
improvements.

1014
00:36:26,360 --> 00:36:28,080
And

1015
00:36:28,080 --> 00:36:30,280
you know, as an an academic, I seen the

1016
00:36:30,280 --> 00:36:32,040
ways of technology that eventually I'm

1017
00:36:32,040 --> 00:36:34,600
taking advantage of as an entrepreneur.

1018
00:36:34,600 --> 00:36:35,840
Um

1019
00:36:35,840 --> 00:36:38,040
I seen them come by and and I I seen

1020
00:36:38,040 --> 00:36:39,840
that they've been largely ignored by

1021
00:36:39,840 --> 00:36:43,120
many of the the larger companies and and

1022
00:36:43,120 --> 00:36:45,320
for that I think we should fault them.

1023
00:36:45,320 --> 00:36:47,640
I I think as engineers,

1024
00:36:47,640 --> 00:36:49,400
you know, and developers, when we have

1025
00:36:49,400 --> 00:36:51,200
the resources to do what needs to be

1026
00:36:51,200 --> 00:36:53,800
done and it's not done, that that's just

1027
00:36:53,800 --> 00:36:56,040
slowing everybody down. And I I think

1028
00:36:56,040 --> 00:36:57,320
there's a little bit of that in the

1029
00:36:57,320 --> 00:37:00,160
legacy vendors to be completely honest.

1030
00:37:00,160 --> 00:37:02,080
One of the things I observed uh I've

1031
00:37:02,080 --> 00:37:03,720
been trying to to see

1032
00:37:03,720 --> 00:37:05,840
if it's going to change or why it's the

1033
00:37:05,840 --> 00:37:07,040
case.

1034
00:37:07,040 --> 00:37:08,440
If you look at

1035
00:37:08,440 --> 00:37:10,440
the automotive sector,

1036
00:37:10,440 --> 00:37:13,720
Mhm. it's largely

1037
00:37:13,720 --> 00:37:15,240
ISV

1038
00:37:15,240 --> 00:37:17,440
or things like OpenFOAM or commercial

1039
00:37:17,440 --> 00:37:19,360
versions of OpenFOAM. Whereas, if you

1040
00:37:19,360 --> 00:37:22,400
look at the aerospace sector, it's a lot

1041
00:37:22,400 --> 00:37:25,320
of homegrown code.

1042
00:37:25,320 --> 00:37:27,040
Why do you think that is and you think

1043
00:37:27,040 --> 00:37:28,480
it will always be that case in the

1044
00:37:28,480 --> 00:37:30,720
aerospace or do you think there's this

1045
00:37:30,720 --> 00:37:32,600
slow

1046
00:37:32,600 --> 00:37:35,400
move and I'm not not because I'm not

1047
00:37:35,400 --> 00:37:36,880
trying to blame you but but because some

1048
00:37:36,880 --> 00:37:38,560
people are maybe going, "You know what?

1049
00:37:38,560 --> 00:37:40,480
I want to have I

1050
00:37:40,480 --> 00:37:42,160
I want to leave that company and go and

1051
00:37:42,160 --> 00:37:44,000
join an exciting startup and do it." Do

1052
00:37:44,000 --> 00:37:46,760
you Do you think eventually there will

1053
00:37:46,760 --> 00:37:49,200
be a move to a more sort of private

1054
00:37:49,200 --> 00:37:52,360
codes and commercial codes and

1055
00:37:52,360 --> 00:37:53,640
So, I

1056
00:37:53,640 --> 00:37:55,840
I witnessed a lot of things through my

1057
00:37:55,840 --> 00:37:57,960
career that lead me to the following

1058
00:37:57,960 --> 00:37:59,280
comments that I'll make about your

1059
00:37:59,280 --> 00:38:01,200
question. So,

1060
00:38:01,200 --> 00:38:02,520
um

1061
00:38:02,520 --> 00:38:04,760
the aerospace industry invested heavily

1062
00:38:04,760 --> 00:38:06,400
in computational fluid dynamics from the

1063
00:38:06,400 --> 00:38:08,440
beginning as an alternative to internal

1064
00:38:08,440 --> 00:38:09,880
testing.

1065
00:38:09,880 --> 00:38:11,560
To the point that the major aerospace

1066
00:38:11,560 --> 00:38:14,560
corporations in the '70s, '80s, and

1067
00:38:14,560 --> 00:38:18,280
'90s, even into the early 2000s,

1068
00:38:18,280 --> 00:38:20,760
would have, you know, teams of 50, 60 to

1069
00:38:20,760 --> 00:38:22,800
100 people working on the development of

1070
00:38:22,800 --> 00:38:24,400
these methods that were used throughout

1071
00:38:24,400 --> 00:38:27,000
their corporations. That changed in the

1072
00:38:27,000 --> 00:38:29,920
early 2000s and the maturity of some of

1073
00:38:29,920 --> 00:38:32,240
the commercial tools became to be so

1074
00:38:32,240 --> 00:38:33,440
high

1075
00:38:33,440 --> 00:38:35,680
that even aerospace companies started

1076
00:38:35,680 --> 00:38:37,760
reducing the size of the teams that were

1077
00:38:37,760 --> 00:38:40,360
doing new method development, new code

1078
00:38:40,360 --> 00:38:42,720
development in favor of sort of

1079
00:38:42,720 --> 00:38:44,840
commercial offerings and options, right?

1080
00:38:44,840 --> 00:38:45,400
Uh

1081
00:38:45,400 --> 00:38:46,600
Um

1082
00:38:46,600 --> 00:38:49,080
that happened more quickly, I think, uh

1083
00:38:49,080 --> 00:38:51,040
particularly the transition to transient

1084
00:38:51,040 --> 00:38:53,760
flow uh calculations in the automotive

1085
00:38:53,760 --> 00:38:55,080
industry.

1086
00:38:55,080 --> 00:38:57,960
They they recognized the advantages of

1087
00:38:57,960 --> 00:38:59,960
the technology. They

1088
00:38:59,960 --> 00:39:01,240
they knew they have tools that were

1089
00:39:01,240 --> 00:39:03,520
almost ready. They some of them invested

1090
00:39:03,520 --> 00:39:05,280
in improving, you know, existing

1091
00:39:05,280 --> 00:39:08,560
solvers, OpenFOAM type uh things, right?

1092
00:39:08,560 --> 00:39:09,120
Uh

1093
00:39:09,120 --> 00:39:10,240
uh there were a number of smaller

1094
00:39:10,240 --> 00:39:11,760
companies that were formed around the

1095
00:39:11,760 --> 00:39:13,640
open-source code, OpenFOAM, in order to

1096
00:39:13,640 --> 00:39:15,240
do further improvements that are needed

1097
00:39:15,240 --> 00:39:17,560
by industry. And largely there's there's

1098
00:39:17,560 --> 00:39:19,440
a lot of penetration, you know, of

1099
00:39:19,440 --> 00:39:21,160
existing commercial vendors and

1100
00:39:21,160 --> 00:39:23,360
open-source-based commercial vendors

1101
00:39:23,360 --> 00:39:25,440
because they seem, you know, the the

1102
00:39:25,440 --> 00:39:27,520
advantages. I

1103
00:39:27,520 --> 00:39:30,360
I I think aerospace industry is going in

1104
00:39:30,360 --> 00:39:31,880
the same direction.

1105
00:39:31,880 --> 00:39:34,400
Um there may be some esoteric

1106
00:39:34,400 --> 00:39:36,200
applications,

1107
00:39:36,200 --> 00:39:38,520
you know, in stealth aircraft and, you

1108
00:39:38,520 --> 00:39:39,960
know, various other things that may

1109
00:39:39,960 --> 00:39:41,520
still be in the realm of what the

1110
00:39:41,520 --> 00:39:43,240
aerospace companies want to do, but the

1111
00:39:43,240 --> 00:39:45,320
vast majority of the simulation

1112
00:39:45,320 --> 00:39:47,440
workflows are ones that commercial

1113
00:39:47,440 --> 00:39:49,080
software can do.

1114
00:39:49,080 --> 00:39:51,400
And that commercial companies that are

1115
00:39:51,400 --> 00:39:53,880
reinventing the way this gets done can

1116
00:39:53,880 --> 00:39:56,560
do way better and way faster. So,

1117
00:39:56,560 --> 00:39:57,960
eventually it's going to go in that

1118
00:39:57,960 --> 00:40:00,520
direction. The the value added by the

1119
00:40:00,520 --> 00:40:02,040
companies is not going to be in the

1120
00:40:02,040 --> 00:40:04,360
development of yet another unstructured

1121
00:40:04,360 --> 00:40:06,080
polyhedral, you know, finite volume

1122
00:40:06,080 --> 00:40:09,040
solver, but rather in how you use it in

1123
00:40:09,040 --> 00:40:10,880
the outer loops. And the inner loops are

1124
00:40:10,880 --> 00:40:12,920
going to be taken for granted if

1125
00:40:12,920 --> 00:40:14,920
companies like ours and others start

1126
00:40:14,920 --> 00:40:16,320
making sure that you can get them

1127
00:40:16,320 --> 00:40:19,120
accurately fast and scalable as as we

1128
00:40:19,120 --> 00:40:21,640
were discussing before. So, So, my take

1129
00:40:21,640 --> 00:40:23,560
is the aerospace industry will go mostly

1130
00:40:23,560 --> 00:40:26,600
commercial, mostly I would say

1131
00:40:26,600 --> 00:40:29,120
modern computing type approaches and

1132
00:40:29,120 --> 00:40:31,120
they'll be building or rebuilding some

1133
00:40:31,120 --> 00:40:34,200
other processes on top of these units of

1134
00:40:34,200 --> 00:40:35,680
computation that are going to be

1135
00:40:35,680 --> 00:40:38,160
effectively taken for granted.

1136
00:40:38,160 --> 00:40:40,840
Yeah, just seem almost ironically that

1137
00:40:40,840 --> 00:40:42,280
it's a

1138
00:40:42,280 --> 00:40:45,640
a credit to CFD if it does become that

1139
00:40:45,640 --> 00:40:47,880
way because almost it feels

1140
00:40:47,880 --> 00:40:50,400
like it had to be developed by your own

1141
00:40:50,400 --> 00:40:52,240
cuz you were the only people who knew

1142
00:40:52,240 --> 00:40:55,000
how to do it where almost now not

1143
00:40:55,000 --> 00:40:56,160
commoditized, but you know, it's

1144
00:40:56,160 --> 00:40:59,280
becoming able to be created and made

1145
00:40:59,280 --> 00:41:01,800
automated that you don't need that to be

1146
00:41:01,800 --> 00:41:03,520
a special team

1147
00:41:03,520 --> 00:41:05,160
within your company.

1148
00:41:05,160 --> 00:41:06,760
Well, in the early days of the aerospace

1149
00:41:06,760 --> 00:41:08,800
industry, Neil, it was a competitive

1150
00:41:08,800 --> 00:41:10,640
advantage to have a team of experts

1151
00:41:10,640 --> 00:41:12,520
developing that capability that nobody

1152
00:41:12,520 --> 00:41:14,000
else had.

1153
00:41:14,000 --> 00:41:16,800
At this point, the the individual

1154
00:41:16,800 --> 00:41:19,560
solution capability is is not something

1155
00:41:19,560 --> 00:41:20,920
that's going to be differentiating

1156
00:41:20,920 --> 00:41:22,840
across these companies, but how you use

1157
00:41:22,840 --> 00:41:25,120
it in tens or hundreds or thousands of

1158
00:41:25,120 --> 00:41:26,480
times, right?

1159
00:41:26,480 --> 00:41:27,720
And then how do you embed it into

1160
00:41:27,720 --> 00:41:29,360
processes and raise other things. So,

1161
00:41:29,360 --> 00:41:31,280
yes, I I think it's it's a credit to the

1162
00:41:31,280 --> 00:41:36,080
success of CFD as a discipline

1163
00:41:36,080 --> 00:41:38,440
that 30, 40, 50 years after it was

1164
00:41:38,440 --> 00:41:40,560
created, it's now becoming something

1165
00:41:40,560 --> 00:41:42,480
that people believe, trust, and they can

1166
00:41:42,480 --> 00:41:44,400
use in that inner loop without giving it

1167
00:41:44,400 --> 00:41:46,800
much thought. And actually relinquishing

1168
00:41:46,800 --> 00:41:48,880
it to to companies like ours, for

1169
00:41:48,880 --> 00:41:51,040
example, and trusting

1170
00:41:51,040 --> 00:41:53,040
that those companies are doing all the

1171
00:41:53,040 --> 00:41:54,920
due diligence enough to make sure that

1172
00:41:54,920 --> 00:41:57,040
the accuracy, performance trade-offs are

1173
00:41:57,040 --> 00:41:59,200
are well understood.

1174
00:41:59,200 --> 00:42:02,280
Yeah, almost feels a way a bit like

1175
00:42:02,280 --> 00:42:03,560
um

1176
00:42:03,560 --> 00:42:07,040
HPC and the cloud. And obviously, I am

1177
00:42:07,040 --> 00:42:09,320
slightly biased. I work for AWS, but I I

1178
00:42:09,320 --> 00:42:11,000
would say this even if I wasn't working.

1179
00:42:11,000 --> 00:42:13,240
And I think others have made these

1180
00:42:13,240 --> 00:42:15,800
comments that it used to be the fact

1181
00:42:15,800 --> 00:42:19,040
that HPC was a very specialist topic,

1182
00:42:19,040 --> 00:42:22,880
required specialist hardware,

1183
00:42:22,880 --> 00:42:24,760
but now

1184
00:42:24,760 --> 00:42:26,400
even if somebody built their own

1185
00:42:26,400 --> 00:42:28,560
machine, they're just using nodes and

1186
00:42:28,560 --> 00:42:31,440
types that are widely used by anybody

1187
00:42:31,440 --> 00:42:33,560
else. They're not really specific, so

1188
00:42:33,560 --> 00:42:36,040
why would you need to build your own

1189
00:42:36,040 --> 00:42:37,840
if you can just get it

1190
00:42:37,840 --> 00:42:40,240
from a cloud vendor. It's because it's

1191
00:42:40,240 --> 00:42:42,360
become more normalized,

1192
00:42:42,360 --> 00:42:43,080
um

1193
00:42:43,080 --> 00:42:45,120
you don't need to have this special

1194
00:42:45,120 --> 00:42:47,520
system in-house.

1195
00:42:47,520 --> 00:42:50,200
I I think you said the keyword before,

1196
00:42:50,200 --> 00:42:51,680
you know, high-performance computing has

1197
00:42:51,680 --> 00:42:53,920
become commoditized.

1198
00:42:53,920 --> 00:42:56,240
You know, no longer do we have

1199
00:42:56,240 --> 00:42:58,360
custom-made chips, custom-made operating

1200
00:42:58,360 --> 00:42:59,920
systems as was the case when I was a

1201
00:42:59,920 --> 00:43:01,640
grad student, you know, custom-made

1202
00:43:01,640 --> 00:43:03,880
interconnects if you have that. This is

1203
00:43:03,880 --> 00:43:06,000
all commoditized. And, you know, the

1204
00:43:06,000 --> 00:43:08,040
cloud service providers are building

1205
00:43:08,040 --> 00:43:10,280
these data centers with the best in

1206
00:43:10,280 --> 00:43:13,960
class of those commoditized components,

1207
00:43:13,960 --> 00:43:15,000
and then they're putting the software

1208
00:43:15,000 --> 00:43:16,720
infrastructure infrastructure together

1209
00:43:16,720 --> 00:43:18,120
so you can actually use them in a

1210
00:43:18,120 --> 00:43:20,160
dynamic way. So,

1211
00:43:20,160 --> 00:43:21,280
um

1212
00:43:21,280 --> 00:43:23,440
it it's hard to believe

1213
00:43:23,440 --> 00:43:26,360
that individual companies are going to

1214
00:43:26,360 --> 00:43:28,960
be able to recreate

1215
00:43:28,960 --> 00:43:30,840
the value that's being added by these

1216
00:43:30,840 --> 00:43:32,680
cloud-based companies at much larger

1217
00:43:32,680 --> 00:43:33,760
scale.

1218
00:43:33,760 --> 00:43:35,360
And, of course, it's hard to believe

1219
00:43:35,360 --> 00:43:37,400
that individual companies are going to

1220
00:43:37,400 --> 00:43:39,280
be able to buy at the scale that's

1221
00:43:39,280 --> 00:43:41,920
needed in order to burst in capacity.

1222
00:43:41,920 --> 00:43:43,520
It's It's hard to believe that

1223
00:43:43,520 --> 00:43:44,920
individual companies are going to be

1224
00:43:44,920 --> 00:43:46,760
able to have access to the GPUs, the

1225
00:43:46,760 --> 00:43:49,440
most modern GPUs, as early as the large

1226
00:43:49,440 --> 00:43:50,880
cloud service providers are having

1227
00:43:50,880 --> 00:43:53,160
access to them. So, and and it's hard to

1228
00:43:53,160 --> 00:43:54,880
believe that

1229
00:43:54,880 --> 00:43:58,200
that companies are going to reproduce

1230
00:43:58,200 --> 00:44:00,880
all of the software services

1231
00:44:00,880 --> 00:44:03,240
that come from a cloud infrastructure

1232
00:44:03,240 --> 00:44:04,600
that that are not there in the

1233
00:44:04,600 --> 00:44:07,000
on-premises clusters. So, yeah, I I

1234
00:44:07,000 --> 00:44:09,320
mean, the the only drawback of at some

1235
00:44:09,320 --> 00:44:12,600
point or drawbacks were the cost and you

1236
00:44:12,600 --> 00:44:14,920
know, the perception of security.

1237
00:44:14,920 --> 00:44:16,840
I mean, the security one is essentially

1238
00:44:16,840 --> 00:44:19,120
gone away and the costs are coming down

1239
00:44:19,120 --> 00:44:21,000
over time. So, different companies have

1240
00:44:21,000 --> 00:44:23,160
different appetites,

1241
00:44:23,160 --> 00:44:24,800
you know, for understanding how much

1242
00:44:24,800 --> 00:44:26,880
they spend internally to achieve similar

1243
00:44:26,880 --> 00:44:28,840
or lower levels of of, you know,

1244
00:44:28,840 --> 00:44:31,320
reliability, you know, credibility,

1245
00:44:31,320 --> 00:44:34,160
security, etc., etc. And and different

1246
00:44:34,160 --> 00:44:36,680
people will jump into the bandwagon at

1247
00:44:36,680 --> 00:44:39,280
steps. But as you have seen, the

1248
00:44:39,280 --> 00:44:41,400
engineering profession was lagging

1249
00:44:41,400 --> 00:44:44,200
behind sort of the financial systems

1250
00:44:44,200 --> 00:44:46,160
that sort of migrated to the cloud 5, 6,

1251
00:44:46,160 --> 00:44:49,040
7 years ago completely and engineering

1252
00:44:49,040 --> 00:44:50,960
is going in that direction right now.

1253
00:44:50,960 --> 00:44:51,520
It's happening.

1254
00:44:51,520 --> 00:44:52,800
Yeah.

1255
00:44:52,800 --> 00:44:56,280
Yeah, it's it seems to be a um

1256
00:44:56,280 --> 00:44:59,320
an interesting transition point where

1257
00:44:59,320 --> 00:45:02,240
there's still I guess like anybody, you

1258
00:45:02,240 --> 00:45:03,840
know, if you say, "Hey, use this new

1259
00:45:03,840 --> 00:45:05,440
turbulence model." No, no, "No, I have to

1260
00:45:05,440 --> 00:45:07,480
use that one." It's it's everybody gets

1261
00:45:07,480 --> 00:45:09,600
in a certain mindset of doing it the way

1262
00:45:09,600 --> 00:45:11,400
they've done it forever.

1263
00:45:11,400 --> 00:45:12,440
Um

1264
00:45:12,440 --> 00:45:14,600
So, it it is hard

1265
00:45:14,600 --> 00:45:17,520
to sometimes convince and even in my

1266
00:45:17,520 --> 00:45:20,560
4 years as it is now with cloud, what it

1267
00:45:20,560 --> 00:45:22,840
was at the beginning to what it is now

1268
00:45:22,840 --> 00:45:24,320
is already a major difference. The

1269
00:45:24,320 --> 00:45:26,240
conversations I was having at the

1270
00:45:26,240 --> 00:45:28,480
beginning, it was being told to sort of

1271
00:45:28,480 --> 00:45:30,480
get lost and I had to

1272
00:45:30,480 --> 00:45:32,040
try, you know, and I'm not a sort of

1273
00:45:32,040 --> 00:45:34,120
salesperson in that sense. Like, I pride

1274
00:45:34,120 --> 00:45:37,040
myself in not selling stuff to people,

1275
00:45:37,040 --> 00:45:39,760
but it it was quite a different People

1276
00:45:39,760 --> 00:45:42,240
didn't I Well, although, would you not

1277
00:45:42,240 --> 00:45:44,720
say that

1278
00:45:44,720 --> 00:45:47,720
you and I in some way

1279
00:45:47,720 --> 00:45:50,480
working in a more cloud or closer to the

1280
00:45:50,480 --> 00:45:51,960
cloud companies, working at a cloud

1281
00:45:51,960 --> 00:45:54,280
company, your company that uses cloud

1282
00:45:54,280 --> 00:45:55,640
resources,

1283
00:45:55,640 --> 00:45:58,600
I still wonder if we're a little bit

1284
00:45:58,600 --> 00:46:00,680
um

1285
00:46:00,680 --> 00:46:03,000
not fully aware of still how much people

1286
00:46:03,000 --> 00:46:05,000
don't even know what the cloud is. You

1287
00:46:05,000 --> 00:46:06,880
know, if you go to some companies, you

1288
00:46:06,880 --> 00:46:08,320
know, like

1289
00:46:08,320 --> 00:46:09,760
does that surprise you still that when

1290
00:46:09,760 --> 00:46:12,640
you're like, "Oh, do you not realize how

1291
00:46:12,640 --> 00:46:13,560
I know it sounds like I'm selling it

1292
00:46:13,560 --> 00:46:15,120
now, but you know how good it is or how

1293
00:46:15,120 --> 00:46:17,600
much potential there is?"

1294
00:46:17,600 --> 00:46:18,880
Some people don't even know what it is,

1295
00:46:18,880 --> 00:46:20,000
yeah.

1296
00:46:20,000 --> 00:46:21,840
I'm not surprised

1297
00:46:21,840 --> 00:46:23,000
that

1298
00:46:23,000 --> 00:46:25,240
people have

1299
00:46:25,240 --> 00:46:28,160
a lack of understanding of the potential

1300
00:46:28,160 --> 00:46:29,720
of the cloud for their engineering

1301
00:46:29,720 --> 00:46:32,760
simulations because there's so many ways

1302
00:46:32,760 --> 00:46:35,440
of using the cloud and I think it's

1303
00:46:35,440 --> 00:46:37,160
cluttering

1304
00:46:37,160 --> 00:46:39,720
their perception of what the cloud does.

1305
00:46:39,720 --> 00:46:41,440
I mean, I I talked to a number of people

1306
00:46:41,440 --> 00:46:42,880
over the last 4 years while we were

1307
00:46:42,880 --> 00:46:44,320
building Luminary

1308
00:46:44,320 --> 00:46:47,760
who really have very little idea

1309
00:46:47,760 --> 00:46:50,000
of what we were saying we were doing at

1310
00:46:50,000 --> 00:46:52,160
Luminary cuz they Again, they had a

1311
00:46:52,160 --> 00:46:53,360
mixture of ideas. They're like, "You

1312
00:46:53,360 --> 00:46:55,440
mean, you're hosting my data there and I

1313
00:46:55,440 --> 00:46:58,320
never have to have something here or is

1314
00:46:58,320 --> 00:47:01,680
it a virtual private cloud or am I just

1315
00:47:01,680 --> 00:47:03,240
using the cloud for when I don't have

1316
00:47:03,240 --> 00:47:04,920
resources in my company?" So, there's

1317
00:47:04,920 --> 00:47:07,280
many ways in which people have done it.

1318
00:47:07,280 --> 00:47:09,040
I I have to tell you, Neil, that the

1319
00:47:09,040 --> 00:47:10,920
best way

1320
00:47:10,920 --> 00:47:13,000
to make sure that people understand how

1321
00:47:13,000 --> 00:47:15,280
we think the cloud should be used for

1322
00:47:15,280 --> 00:47:17,320
engineering simulations is to show them

1323
00:47:17,320 --> 00:47:18,520
a demo.

1324
00:47:18,520 --> 00:47:21,160
Oh, yeah. Right. Yeah. It It's when it

1325
00:47:21,160 --> 00:47:23,080
clicks for people. I can see that all

1326
00:47:23,080 --> 00:47:24,480
the time. They have all these questions

1327
00:47:24,480 --> 00:47:25,800
in their minds of, "What do you mean

1328
00:47:25,800 --> 00:47:28,120
this, that, that?" When you finally show

1329
00:47:28,120 --> 00:47:30,520
a demo and you upload a CAD file to the

1330
00:47:30,520 --> 00:47:33,960
cloud and it takes, you know, 2 seconds

1331
00:47:33,960 --> 00:47:36,320
and then you run a transient simulation

1332
00:47:36,320 --> 00:47:38,480
that generates multiple terabytes and

1333
00:47:38,480 --> 00:47:40,440
immediately you can actually see and

1334
00:47:40,440 --> 00:47:42,360
you've never transferred a single file

1335
00:47:42,360 --> 00:47:44,800
or done anything, that's when people see

1336
00:47:44,800 --> 00:47:47,720
it and and I I think it's it's companies

1337
00:47:47,720 --> 00:47:49,720
like ours and others who are beginning

1338
00:47:49,720 --> 00:47:51,080
to do this

1339
00:47:51,080 --> 00:47:53,160
that can give people the more clear

1340
00:47:53,160 --> 00:47:55,600
understanding of what it is and what the

1341
00:47:55,600 --> 00:47:58,880
potential it has of becoming

1342
00:47:58,880 --> 00:48:00,560
and and that's when it clicks for people

1343
00:48:00,560 --> 00:48:03,000
I think normally. Mhm.

1344
00:48:03,000 --> 00:48:04,880
But yes, there's many different ways as

1345
00:48:04,880 --> 00:48:07,600
of you know of using the cloud and and I

1346
00:48:07,600 --> 00:48:09,520
think that's confused a lot of people

1347
00:48:09,520 --> 00:48:11,480
who who've been around simulation over

1348
00:48:11,480 --> 00:48:12,760
the years but have not paid close

1349
00:48:12,760 --> 00:48:14,440
attention to the cloud.

1350
00:48:14,440 --> 00:48:17,040
Well, I'm I'm kind of um

1351
00:48:17,040 --> 00:48:20,320
glad in a way that um it has turned out

1352
00:48:20,320 --> 00:48:23,160
the way it does because I remember I

1353
00:48:23,160 --> 00:48:24,280
was giving a

1354
00:48:24,280 --> 00:48:25,600
talk.

1355
00:48:25,600 --> 00:48:28,880
I think it was in about May of 2020,

1356
00:48:28,880 --> 00:48:30,560
something like that. I think it was like

1357
00:48:30,560 --> 00:48:33,840
a remote thing at um at NASA. They do

1358
00:48:33,840 --> 00:48:36,160
this seminar series and I gave a talk on

1359
00:48:36,160 --> 00:48:38,480
like how the cloud

1360
00:48:38,480 --> 00:48:40,800
you know, will transform CFD and I I

1361
00:48:40,800 --> 00:48:43,400
remember the I mean admittedly okay, for

1362
00:48:43,400 --> 00:48:45,240
people who know I was doing at NASA's

1363
00:48:45,240 --> 00:48:46,840
supercomputing division which probably

1364
00:48:46,840 --> 00:48:48,920
is not the right place to be pitching.

1365
00:48:48,920 --> 00:48:49,960
Sure.

1366
00:48:49,960 --> 00:48:52,120
Sure.

1367
00:48:52,120 --> 00:48:54,040
Um but anyway, I remember saying some of

1368
00:48:54,040 --> 00:48:55,320
the stuff and I

1369
00:48:55,320 --> 00:48:56,920
I genuinely thought it was going to be

1370
00:48:56,920 --> 00:49:00,160
the case but um there was not as many

1371
00:49:00,160 --> 00:49:02,440
other people truly believing in it and

1372
00:49:02,440 --> 00:49:04,520
it was interesting that

1373
00:49:04,520 --> 00:49:05,920
in that

1374
00:49:05,920 --> 00:49:09,440
even before that your company was

1375
00:49:09,440 --> 00:49:10,960
brewing and it it it showed that it

1376
00:49:10,960 --> 00:49:13,120
takes time for things to happen but it

1377
00:49:13,120 --> 00:49:14,560
I kind of

1378
00:49:14,560 --> 00:49:16,480
at least I'm happy that it it kind of

1379
00:49:16,480 --> 00:49:18,520
has I wasn't telling people a bunch of

1380
00:49:18,520 --> 00:49:20,480
lies and

1381
00:49:20,480 --> 00:49:22,160
has kind of come to be true.

1382
00:49:22,160 --> 00:49:24,280
You were one of the pioneers of this. I

1383
00:49:24,280 --> 00:49:26,080
remember you and I we met at

1384
00:49:26,080 --> 00:49:28,280
I think it was an international CFD

1385
00:49:28,280 --> 00:49:31,040
conference in Strathclyde.

1386
00:49:31,040 --> 00:49:33,200
Oh, yeah. A few years few years before

1387
00:49:33,200 --> 00:49:36,000
that, maybe in 2016-17.

1388
00:49:36,000 --> 00:49:37,520
And you were beginning to talk about

1389
00:49:37,520 --> 00:49:39,680
these things and and I was paying

1390
00:49:39,680 --> 00:49:41,120
attention.

1391
00:49:41,120 --> 00:49:44,160
And yes, I I think

1392
00:49:44,160 --> 00:49:45,360
you know, going to people who own

1393
00:49:45,360 --> 00:49:47,520
supercomputers and that are very

1394
00:49:47,520 --> 00:49:48,880
interested in sort of the hero

1395
00:49:48,880 --> 00:49:50,320
calculations where you're going to be

1396
00:49:50,320 --> 00:49:53,880
using 1 and 1/2 million, you know, CPUs

1397
00:49:53,880 --> 00:49:56,160
or something. It It That's the right the

1398
00:49:56,160 --> 00:49:58,320
wrong crowd, right? Uh

1399
00:49:58,320 --> 00:50:00,000
But But as you very well know, you know,

1400
00:50:00,000 --> 00:50:02,800
from the early days in the early '90s of

1401
00:50:02,800 --> 00:50:04,200
MPI,

1402
00:50:04,200 --> 00:50:07,480
you know, 4 6 8 16 processors,

1403
00:50:07,480 --> 00:50:10,280
eventually 200 400, then Blue Gene L

1404
00:50:10,280 --> 00:50:13,240
with 100,000 cores or racks, and then

1405
00:50:13,240 --> 00:50:15,000
eventually some of the 1.5 million

1406
00:50:15,000 --> 00:50:17,360
simulations. We all thought that for

1407
00:50:17,360 --> 00:50:19,560
supercomputing to keep going, we're

1408
00:50:19,560 --> 00:50:21,920
going to get into the millions of cores,

1409
00:50:21,920 --> 00:50:24,360
yeah, yeah, heterogeneous, etc., etc.,

1410
00:50:24,360 --> 00:50:25,520
right?

1411
00:50:25,520 --> 00:50:27,080
Uh and everybody was talking about MPI

1412
00:50:27,080 --> 00:50:30,240
plus X. We didn't know how the future of

1413
00:50:30,240 --> 00:50:31,680
software development was going to

1414
00:50:31,680 --> 00:50:34,040
happen. It turns out that for some of

1415
00:50:34,040 --> 00:50:36,240
the most demanding supercomputing

1416
00:50:36,240 --> 00:50:38,280
calculations for engineering, I'm not

1417
00:50:38,280 --> 00:50:39,960
going to talk about basic science at the

1418
00:50:39,960 --> 00:50:42,400
moment, but for engineering,

1419
00:50:42,400 --> 00:50:45,720
we can do really well with 256 A100 GPUs

1420
00:50:45,720 --> 00:50:47,400
in the cloud. So,

1421
00:50:47,400 --> 00:50:49,920
so I think it's taking a while for

1422
00:50:49,920 --> 00:50:52,080
people to understand

1423
00:50:52,080 --> 00:50:54,960
that vast improvements in supercomputing

1424
00:50:54,960 --> 00:50:57,800
are not requiring some massive changes,

1425
00:50:57,800 --> 00:50:59,440
you know. It's cloud, yes, new

1426
00:50:59,440 --> 00:51:02,520
technology. It is, you know,

1427
00:51:02,520 --> 00:51:05,920
500 GPUs, 1,000 GPUs, but not hundreds

1428
00:51:05,920 --> 00:51:08,640
of thousands or millions of GPUs, right?

1429
00:51:08,640 --> 00:51:11,120
So, so I I think we got into a point

1430
00:51:11,120 --> 00:51:13,720
where where most people are seeing that

1431
00:51:13,720 --> 00:51:15,360
for engineering calculations, this is

1432
00:51:15,360 --> 00:51:16,920
the way to go.

1433
00:51:16,920 --> 00:51:18,720
So, where I I agree with you there. I

1434
00:51:18,720 --> 00:51:21,440
think that's an interesting

1435
00:51:21,440 --> 00:51:23,480
I always do this. I always show this in

1436
00:51:23,480 --> 00:51:25,320
any presentation I do about like the

1437
00:51:25,320 --> 00:51:26,760
cost

1438
00:51:26,760 --> 00:51:28,840
and that doing an, you know, a DNS or

1439
00:51:28,840 --> 00:51:31,160
things like that are are so unbelievably

1440
00:51:31,160 --> 00:51:35,520
expensive that it's it's good maybe for

1441
00:51:35,520 --> 00:51:37,680
academic work or things where for

1442
00:51:37,680 --> 00:51:39,200
national security reasons or whatever,

1443
00:51:39,200 --> 00:51:41,000
you know, you have to get the answer

1444
00:51:41,000 --> 00:51:42,680
through CFD and

1445
00:51:42,680 --> 00:51:44,440
maybe the accuracy. But I I would agree

1446
00:51:44,440 --> 00:51:47,120
with you that we've made very large

1447
00:51:47,120 --> 00:51:49,960
inroads into showing high fidelity

1448
00:51:49,960 --> 00:51:51,880
methods are maybe not as expensive as

1449
00:51:51,880 --> 00:51:53,760
people thought they were maybe 10 years

1450
00:51:53,760 --> 00:51:55,600
ago. You know, that sort of vision of

1451
00:51:55,600 --> 00:51:57,160
2030

1452
00:51:57,160 --> 00:51:59,120
I would argue is come closer and

1453
00:51:59,120 --> 00:52:00,680
obviously, you know, that was something

1454
00:52:00,680 --> 00:52:02,520
you were heavily involved with. Do you

1455
00:52:02,520 --> 00:52:04,040
think

1456
00:52:04,040 --> 00:52:05,640
How do you think things are looking? So

1457
00:52:05,640 --> 00:52:09,120
the CFD Vision 2030 report was a very big

1458
00:52:09,120 --> 00:52:12,120
pioneering is quoted everywhere. How are

1459
00:52:12,120 --> 00:52:13,880
things progressing towards achieving

1460
00:52:13,880 --> 00:52:14,920
those

1461
00:52:14,920 --> 00:52:15,520
aims?

1462
00:52:15,520 --> 00:52:17,520
Yeah.

1463
00:52:17,560 --> 00:52:19,040
So as you mentioned I was one of the

1464
00:52:19,040 --> 00:52:20,640
co-authors of that study and it was a

1465
00:52:20,640 --> 00:52:22,480
very fun study to do cuz it involved

1466
00:52:22,480 --> 00:52:24,560
academics, it involved industry, you

1467
00:52:24,560 --> 00:52:26,600
know, it involved people who have been

1468
00:52:26,600 --> 00:52:28,160
in the computer science side of the

1469
00:52:28,160 --> 00:52:30,400
effort etc. etc.

1470
00:52:30,400 --> 00:52:31,880
At the time we were so that was

1471
00:52:31,880 --> 00:52:33,840
published in 2014

1472
00:52:33,840 --> 00:52:35,320
10 years ago.

1473
00:52:35,320 --> 00:52:36,640
We've been working on it for about 2

1474
00:52:36,640 --> 00:52:38,640
years. I think is the total time that it

1475
00:52:38,640 --> 00:52:41,160
took us a a small contract from NASA to

1476
00:52:41,160 --> 00:52:43,360
sort of put this vision together of what

1477
00:52:43,360 --> 00:52:47,320
CFD in the 2030 should actually be. Um

1478
00:52:47,320 --> 00:52:50,240
at the time that we were writing that

1479
00:52:50,240 --> 00:52:53,520
report and we published it

1480
00:52:53,520 --> 00:52:56,120
I don't think anyone of us of 10 or so

1481
00:52:56,120 --> 00:52:57,720
people in the committee that wrote that

1482
00:52:57,720 --> 00:52:59,920
paper thought we would get anywhere

1483
00:52:59,920 --> 00:53:02,800
close to being at the level predicted by

1484
00:53:02,800 --> 00:53:07,240
2030. And here we are 6 years before

1485
00:53:07,240 --> 00:53:08,800
the deadline

1486
00:53:08,800 --> 00:53:11,240
and I think much of what we were talking

1487
00:53:11,240 --> 00:53:14,240
about in the solver technology

1488
00:53:14,240 --> 00:53:15,880
in the managing large numbers of

1489
00:53:15,880 --> 00:53:17,880
simulations, in some of the

1490
00:53:17,880 --> 00:53:20,360
multiphysics elements, and some of the

1491
00:53:20,360 --> 00:53:22,800
design optimization elements that I'm

1492
00:53:22,800 --> 00:53:24,320
you know, I that was one of the major

1493
00:53:24,320 --> 00:53:27,240
writers for, I think we're almost there.

1494
00:53:27,240 --> 00:53:28,760
Um

1495
00:53:28,760 --> 00:53:30,200
meshing

1496
00:53:30,200 --> 00:53:33,400
adaptation, accuracy, scale resolving

1497
00:53:33,400 --> 00:53:35,280
simulations, those were things that were

1498
00:53:35,280 --> 00:53:37,720
not deemed to be possible

1499
00:53:37,720 --> 00:53:41,120
you know, until 2030 or beyond and and I

1500
00:53:41,120 --> 00:53:42,960
I think I think we're going to

1501
00:53:42,960 --> 00:53:45,000
definitely meet the goals by 2030 or

1502
00:53:45,000 --> 00:53:46,520
before.

1503
00:53:46,520 --> 00:53:48,440
Which by the way is a lesson for the

1504
00:53:48,440 --> 00:53:50,560
next 20 years, right? The Like I was

1505
00:53:50,560 --> 00:53:52,200
telling you at Stanford people were

1506
00:53:52,200 --> 00:53:54,800
looking at how to program in GPUs before

1507
00:53:54,800 --> 00:53:57,160
there was CUDA or anything else 20 years

1508
00:53:57,160 --> 00:54:00,560
ago. So, you know, we're in 2024, 2044,

1509
00:54:00,560 --> 00:54:03,200
hopefully we're both still around.

1510
00:54:03,200 --> 00:54:05,720
And it could be radically different than

1511
00:54:05,720 --> 00:54:08,240
than what people are working on today

1512
00:54:08,240 --> 00:54:10,080
may actually have a substantial impact

1513
00:54:10,080 --> 00:54:12,640
when, right? So, so we we we're

1514
00:54:12,640 --> 00:54:15,040
definitely on track to achieving the

1515
00:54:15,040 --> 00:54:18,400
goals of the CFD Vision 2030 uh CFD vision

1516
00:54:18,400 --> 00:54:19,680
before.

1517
00:54:19,680 --> 00:54:20,840
Uh there's a group that's actually

1518
00:54:20,840 --> 00:54:22,440
tracking this and and you could argue

1519
00:54:22,440 --> 00:54:23,840
that you could interpret all the

1520
00:54:23,840 --> 00:54:25,400
statements made in the original report

1521
00:54:25,400 --> 00:54:27,680
in one way or another, but but we made

1522
00:54:27,680 --> 00:54:30,520
tremendous progress mostly thanks to GPU

1523
00:54:30,520 --> 00:54:33,040
computing and sort of advances in sort

1524
00:54:33,040 --> 00:54:35,960
of meshing, adaptation, post-processing,

1525
00:54:35,960 --> 00:54:38,760
and multiphysics, right? I think

1526
00:54:38,760 --> 00:54:40,400
I think taking calculations that used to

1527
00:54:40,400 --> 00:54:42,440
take 4 to 6 hours and doing them in 2

1528
00:54:42,440 --> 00:54:45,160
minutes opens up the possibilities for

1529
00:54:45,160 --> 00:54:47,240
multiphysics which were recognized in

1530
00:54:47,240 --> 00:54:49,280
the CFD Vision 2030 as one of the key

1531
00:54:49,280 --> 00:54:52,280
ingredients, you know, for for future

1532
00:54:52,280 --> 00:54:55,360
uses of CFD in in you know, beyond

1533
00:54:55,360 --> 00:54:56,880
what's been used today.

1534
00:54:56,880 --> 00:55:00,880
Mhm. I think one um and I probably

1535
00:55:00,880 --> 00:55:03,320
should have read the report before

1536
00:55:03,320 --> 00:55:04,920
saying this, but I'm pretty sure it

1537
00:55:04,920 --> 00:55:06,560
doesn't have it in there, but correct me

1538
00:55:06,560 --> 00:55:08,320
if I'm wrong.

1539
00:55:08,320 --> 00:55:10,240
Is as we move into now the machine

1540
00:55:10,240 --> 00:55:11,360
learning

1541
00:55:11,360 --> 00:55:13,960
side of things. Um

1542
00:55:13,960 --> 00:55:15,920
I'm interested to know your thoughts on

1543
00:55:15,920 --> 00:55:17,760
this cuz just as

1544
00:55:17,760 --> 00:55:20,440
GPUs, totally agree, have been a

1545
00:55:20,440 --> 00:55:23,840
game-changer, rewriting codes um for

1546
00:55:23,840 --> 00:55:26,840
GPUs, and you know, all the advantages

1547
00:55:26,840 --> 00:55:31,240
that gave and and and and and even just

1548
00:55:31,240 --> 00:55:33,960
as you were saying, integrating the HPC

1549
00:55:33,960 --> 00:55:35,840
and the software more tightly,

1550
00:55:35,840 --> 00:55:40,040
um so it's not separate things.

1551
00:55:40,040 --> 00:55:40,960
But

1552
00:55:40,960 --> 00:55:43,280
where do you see machine learning

1553
00:55:43,280 --> 00:55:46,520
coming? Do you think it will be a major

1554
00:55:46,520 --> 00:55:50,640
change or do you think it's too hyped?

1555
00:55:50,640 --> 00:55:54,040
Where's your viewpoint on this?

1556
00:55:54,040 --> 00:55:55,560
Yeah, so

1557
00:55:55,560 --> 00:55:58,160
this is a long topic and that one that's

1558
00:55:58,160 --> 00:55:59,320
close and dear to my heart. I've been

1559
00:55:59,320 --> 00:56:01,280
doing research in this at Stanford since

1560
00:56:01,280 --> 00:56:03,840
I would say 2010 or so.

1561
00:56:03,840 --> 00:56:05,880
Um

1562
00:56:05,880 --> 00:56:08,200
There is no question

1563
00:56:08,200 --> 00:56:10,360
that it's it's one of our technological

1564
00:56:10,360 --> 00:56:11,800
revolutions

1565
00:56:11,800 --> 00:56:13,560
in general.

1566
00:56:13,560 --> 00:56:15,640
In science, engineering, and consumer

1567
00:56:15,640 --> 00:56:18,120
industries, etc., etc.

1568
00:56:18,120 --> 00:56:21,120
The abundance of data is what motivated

1569
00:56:21,120 --> 00:56:23,360
this, whether you produce it through

1570
00:56:23,360 --> 00:56:25,960
more accurate

1571
00:56:25,960 --> 00:56:27,760
much quicker simulations that you can

1572
00:56:27,760 --> 00:56:29,400
run in parallel, you know, hundreds of

1573
00:56:29,400 --> 00:56:30,800
them, etc., etc., or whether you're

1574
00:56:30,800 --> 00:56:32,760
collecting data from sensors, sensor

1575
00:56:32,760 --> 00:56:35,640
technology from existing systems feel it

1576
00:56:35,640 --> 00:56:36,960
in the

1577
00:56:36,960 --> 00:56:38,120
you know,

1578
00:56:38,120 --> 00:56:39,680
in the world.

1579
00:56:39,680 --> 00:56:42,560
Um, I do think it's a revolution, but I

1580
00:56:42,560 --> 00:56:44,320
am afraid it is

1581
00:56:44,320 --> 00:56:47,400
massively hyped up when it comes to

1582
00:56:47,400 --> 00:56:50,920
physics-based simulation impact of AI

1583
00:56:50,920 --> 00:56:52,960
and ML. And actually, if you will allow

1584
00:56:52,960 --> 00:56:55,880
me, I'll always say ML and AI.

1585
00:56:55,880 --> 00:56:57,920
Because I mean, what we're talking about

1586
00:56:57,920 --> 00:56:59,640
is is

1587
00:56:59,640 --> 00:57:02,160
advanced methodologies

1588
00:57:02,160 --> 00:57:05,120
for regression of existing data,

1589
00:57:05,120 --> 00:57:07,600
mathematically speaking, right?

1590
00:57:07,600 --> 00:57:09,800
Um

1591
00:57:09,800 --> 00:57:11,960
I don't think it's a question of whether

1592
00:57:11,960 --> 00:57:14,320
it will be useful or not. It will be

1593
00:57:14,320 --> 00:57:17,560
useful. I think the key question is for

1594
00:57:17,560 --> 00:57:20,040
what will it be useful?

1595
00:57:20,040 --> 00:57:23,200
And also, how will it be useful for

1596
00:57:23,200 --> 00:57:25,680
people who are doing simulation-based

1597
00:57:25,680 --> 00:57:27,480
work, right? So, I got I got I have to

1598
00:57:27,480 --> 00:57:29,240
carve out that niche, which is my niche,

1599
00:57:29,240 --> 00:57:30,640
right?

1600
00:57:30,640 --> 00:57:33,080
Um, it is my strong opinion at the

1601
00:57:33,080 --> 00:57:34,800
moment and I've been thinking about this

1602
00:57:34,800 --> 00:57:36,240
quite a bit that there are four

1603
00:57:36,240 --> 00:57:40,000
potential uses of ML and AI in

1604
00:57:40,000 --> 00:57:41,680
simulation-driven

1605
00:57:41,680 --> 00:57:45,000
workflows. By the way, if I hear anybody

1606
00:57:45,000 --> 00:57:47,320
put a LinkedIn post or publish something

1607
00:57:47,320 --> 00:57:49,880
when they say like, "Ooh, we did this

1608
00:57:49,880 --> 00:57:51,880
and it's 10,000 times faster than the

1609
00:57:51,880 --> 00:57:53,840
original solver."

1610
00:57:53,840 --> 00:57:56,680
I'm going to write a nasty gram because

1611
00:57:56,680 --> 00:57:58,480
they never talk about how much it took

1612
00:57:58,480 --> 00:58:00,640
to generate the data, how much it took

1613
00:58:00,640 --> 00:58:01,840
to train the model.

1614
00:58:01,840 --> 00:58:03,720
Something recently on

1615
00:58:03,720 --> 00:58:05,640
I wrote a comment on a few days ago. I

1616
00:58:05,640 --> 00:58:07,040
didn't say it was 10,000, to be clear.

1617
00:58:07,040 --> 00:58:09,720
Somebody else did. And I Okay.

1618
00:58:09,720 --> 00:58:11,560
Well, so what what I'm trying to say is

1619
00:58:11,560 --> 00:58:14,800
that those things are are hype.

1620
00:58:14,800 --> 00:58:17,280
No discussion of how many actual

1621
00:58:17,280 --> 00:58:18,800
simulations were required to train the

1622
00:58:18,800 --> 00:58:19,800
model,

1623
00:58:19,800 --> 00:58:21,160
right?

1624
00:58:21,160 --> 00:58:22,960
No discussion

1625
00:58:22,960 --> 00:58:25,760
as to what the training costs were.

1626
00:58:25,760 --> 00:58:28,200
No discussion as to the accuracy of the

1627
00:58:28,200 --> 00:58:30,840
predictive tool. No discussion as to

1628
00:58:30,840 --> 00:58:33,520
whether that methodology is

1629
00:58:33,520 --> 00:58:35,400
interpolative in nature only or

1630
00:58:35,400 --> 00:58:37,840
extrapolative in nature, typically not,

1631
00:58:37,840 --> 00:58:39,120
obviously.

1632
00:58:39,120 --> 00:58:42,520
And you know, sort of no discussion as

1633
00:58:42,520 --> 00:58:44,720
to whether, you know, you were actually

1634
00:58:44,720 --> 00:58:46,760
training for something and actually

1635
00:58:46,760 --> 00:58:49,880
testing for something else. So So my

1636
00:58:49,880 --> 00:58:52,240
take is that those things just ring very

1637
00:58:52,240 --> 00:58:53,560
hollow.

1638
00:58:53,560 --> 00:58:55,400
Most people do not understand the real

1639
00:58:55,400 --> 00:58:57,800
problems for which these replacements of

1640
00:58:57,800 --> 00:58:59,640
high-performance computing sort of type

1641
00:58:59,640 --> 00:59:01,160
of approaches would actually be used

1642
00:59:01,160 --> 00:59:03,200
for. And I I hope I'm not insulting you

1643
00:59:03,200 --> 00:59:05,720
or any of your of your

1644
00:59:05,720 --> 00:59:08,880
you know, podcast viewers and listeners.

1645
00:59:08,880 --> 00:59:09,680
Um

1646
00:59:09,680 --> 00:59:13,280
I I believe strongly that there are uses

1647
00:59:13,280 --> 00:59:15,040
for it, but I think we have to sort of

1648
00:59:15,040 --> 00:59:16,760
get over the hype and and sort of move

1649
00:59:16,760 --> 00:59:19,240
on to to show things. So So if I can

1650
00:59:19,240 --> 00:59:20,720
tell you

1651
00:59:20,720 --> 00:59:23,400
um the four things where I think uh this

1652
00:59:23,400 --> 00:59:26,000
technology is going to be very useful.

1653
00:59:26,000 --> 00:59:26,880
Um

1654
00:59:26,880 --> 00:59:29,080
I I Do you have time? Yeah. Yeah,

1655
00:59:29,080 --> 00:59:31,760
absolutely. Let me start. Number one,

1656
00:59:31,760 --> 00:59:34,520
one that makes absolute sense to me in

1657
00:59:34,520 --> 00:59:36,840
aerospace and outside of aerospace,

1658
00:59:36,840 --> 00:59:40,200
real-time control of systems.

1659
00:59:40,200 --> 00:59:42,120
You have a system, a plan that's fixed.

1660
00:59:42,120 --> 00:59:44,200
You can analyze and hyper-analyze it and

1661
00:59:44,200 --> 00:59:46,120
create training data,

1662
00:59:46,120 --> 00:59:47,480
right? The

1663
00:59:47,480 --> 00:59:50,080
it requires that you query the model

1664
00:59:50,080 --> 00:59:52,240
many times per second. So, there's a

1665
00:59:52,240 --> 00:59:54,040
performance requirement that is

1666
00:59:54,040 --> 00:59:55,880
absolutely necessary.

1667
00:59:55,880 --> 00:59:58,080
It is a control system and therefore

1668
00:59:58,080 --> 01:00:00,760
it's built to reject errors,

1669
01:00:00,760 --> 01:00:02,760
which means it's okay

1670
01:00:02,760 --> 01:00:05,160
to actually build a model that has some

1671
01:00:05,160 --> 01:00:07,520
errors. And in fact, it's okay to build

1672
01:00:07,520 --> 01:00:09,720
models that have significant errors, so

1673
01:00:09,720 --> 01:00:11,240
you can reduce the amount of training

1674
01:00:11,240 --> 01:00:12,920
data, right? The

1675
01:00:12,920 --> 01:00:14,680
And at the end of the day, like I said,

1676
01:00:14,680 --> 01:00:17,080
the system's not changing, so so you can

1677
01:00:17,080 --> 01:00:18,800
actually build it and you can amortize

1678
01:00:18,800 --> 01:00:21,280
it over, you know, many many many

1679
01:00:21,280 --> 01:00:23,480
repetitions of that system, whether it's

1680
01:00:23,480 --> 01:00:26,040
an airplane or a jet engine or a car or

1681
01:00:26,040 --> 01:00:28,040
something else. That makes perfect sense

1682
01:00:28,040 --> 01:00:30,200
to me. It's a very viable use. So,

1683
01:00:30,200 --> 01:00:32,040
that's that to me will be use number

1684
01:00:32,040 --> 01:00:33,240
one.

1685
01:00:33,240 --> 01:00:35,360
Uh, use number two,

1686
01:00:35,360 --> 01:00:37,880
uh, would be some very hard optimization

1687
01:00:37,880 --> 01:00:39,640
problems. So, design, design

1688
01:00:39,640 --> 01:00:41,760
optimization for various different

1689
01:00:41,760 --> 01:00:43,280
reasons. Maybe it's very high

1690
01:00:43,280 --> 01:00:44,760
dimensional, right? Maybe there has

1691
01:00:44,760 --> 01:00:46,680
noise. Maybe it's very multiphysics

1692
01:00:46,680 --> 01:00:49,280
oriented, etc., etc. I can see how what

1693
01:00:49,280 --> 01:00:52,520
I would call ML/AI surrogates

1694
01:00:52,520 --> 01:00:54,480
with certain amount of errors in the

1695
01:00:54,480 --> 01:00:56,800
context of an optimization framework,

1696
01:00:56,800 --> 01:00:58,920
like let's say a trust region based type

1697
01:00:58,920 --> 01:01:01,320
of approach, could actually be helpful,

1698
01:01:01,320 --> 01:01:03,120
more helpful than doing direct

1699
01:01:03,120 --> 01:01:06,600
optimization on top of the high fidelity

1700
01:01:06,600 --> 01:01:09,720
analysis. Except in many situations, you

1701
01:01:09,720 --> 01:01:11,120
know, if I can get away with adjoint

1702
01:01:11,120 --> 01:01:12,800
methods and various other sophisticated

1703
01:01:12,800 --> 01:01:15,120
methodologies, well, let's say 100 or

1704
01:01:15,120 --> 01:01:18,120
200 function evaluations, well, and and

1705
01:01:18,120 --> 01:01:19,600
get to the optimum with high fidelity

1706
01:01:19,600 --> 01:01:21,400
without having to check and recheck, why

1707
01:01:21,400 --> 01:01:23,160
am I going to train a model with several

1708
01:01:23,160 --> 01:01:25,360
thousand simulations, you know, unless

1709
01:01:25,360 --> 01:01:27,640
I'm going to be doing it very often. So,

1710
01:01:27,640 --> 01:01:29,600
so some of these optimization hard

1711
01:01:29,600 --> 01:01:32,280
optimization problems could benefit from

1712
01:01:32,280 --> 01:01:33,960
these types of methodologies. That to me

1713
01:01:33,960 --> 01:01:36,400
is application number two, right there.

1714
01:01:36,400 --> 01:01:38,600
Um application number three are some of

1715
01:01:38,600 --> 01:01:40,840
the outer loops that also can tolerate

1716
01:01:40,840 --> 01:01:43,200
some errors. So, uncertainty

1717
01:01:43,200 --> 01:01:44,680
quantification, design under

1718
01:01:44,680 --> 01:01:47,680
uncertainty, some large parameter

1719
01:01:47,680 --> 01:01:50,160
studies, data assimilation, inverse

1720
01:01:50,160 --> 01:01:52,080
problems. There are some situations

1721
01:01:52,080 --> 01:01:53,840
where this makes sense. And then

1722
01:01:53,840 --> 01:01:56,480
finally, if you have a system that is

1723
01:01:56,480 --> 01:01:58,840
not only simulatable

1724
01:01:58,840 --> 01:02:00,920
and trainable, but you can actually

1725
01:02:00,920 --> 01:02:03,360
collect a lot of data. So, digital twin

1726
01:02:03,360 --> 01:02:06,000
type ideas in certain situations. The

1727
01:02:06,000 --> 01:02:07,960
the ability of having frameworks,

1728
01:02:07,960 --> 01:02:09,400
possibly cloud-based, where you're

1729
01:02:09,400 --> 01:02:11,120
constantly retraining based on the

1730
01:02:11,120 --> 01:02:14,080
availability of data for bespoke models

1731
01:02:14,080 --> 01:02:15,320
for various different, you know,

1732
01:02:15,320 --> 01:02:17,240
products that you have out there. That

1733
01:02:17,240 --> 01:02:19,000
could actually be very useful. So, so I

1734
01:02:19,000 --> 01:02:21,320
can see a tremendous potential for those

1735
01:02:21,320 --> 01:02:24,200
four types of applications, but

1736
01:02:24,200 --> 01:02:26,280
they're all always going to be based on

1737
01:02:26,280 --> 01:02:27,760
the high-fidelity simulations that you

1738
01:02:27,760 --> 01:02:30,160
have to create in the first place. And

1739
01:02:30,160 --> 01:02:31,520
they're going to have to be based on

1740
01:02:31,520 --> 01:02:33,360
some more modern, more non-linear

1741
01:02:33,360 --> 01:02:35,640
techniques. And hopefully, as we move

1742
01:02:35,640 --> 01:02:37,840
forward over the next 10 years or so,

1743
01:02:37,840 --> 01:02:40,360
academics, industry, and others will

1744
01:02:40,360 --> 01:02:42,640
begin looking for ways in which these

1745
01:02:42,640 --> 01:02:45,000
models can be much more extrapolative

1746
01:02:45,000 --> 01:02:47,400
than they are today. So, the cost and

1747
01:02:47,400 --> 01:02:49,560
investment of setting up the model and

1748
01:02:49,560 --> 01:02:52,120
training it can be amortized over much

1749
01:02:52,120 --> 01:02:55,720
larger numbers of user uses um than than

1750
01:02:55,720 --> 01:02:57,960
what is the limited set of uses that one

1751
01:02:57,960 --> 01:03:00,520
can use today. So, so I'm not cynical. I

1752
01:03:00,520 --> 01:03:02,880
hope I don't come across as cynical. I I

1753
01:03:02,880 --> 01:03:05,080
I'm more about sort of trying to

1754
01:03:05,080 --> 01:03:07,320
understand from a very rigorous

1755
01:03:07,320 --> 01:03:09,080
mathematical and physical point of view

1756
01:03:09,080 --> 01:03:11,160
what these methodologies can and cannot

1757
01:03:11,160 --> 01:03:12,080
do,

1758
01:03:12,080 --> 01:03:14,640
Yeah. Um so, that that's my thinking.

1759
01:03:14,640 --> 01:03:15,880
There's there's There's of generative

1760
01:03:15,880 --> 01:03:17,880
design that we could go into, that may

1761
01:03:17,880 --> 01:03:20,040
be enabled by the high fidelity

1762
01:03:20,040 --> 01:03:21,960
simulations, but but that's even at an

1763
01:03:21,960 --> 01:03:23,960
earlier stage at the moment. So

1764
01:03:23,960 --> 01:03:25,520
Yeah, I find Yeah, I definitely find an

1765
01:03:25,520 --> 01:03:27,840
interesting topic because um

1766
01:03:27,840 --> 01:03:28,600
Mhm.

1767
01:03:28,600 --> 01:03:30,360
we

1768
01:03:30,360 --> 01:03:32,560
you know, we see a lot of

1769
01:03:32,560 --> 01:03:34,960
interesting companies come out.

1770
01:03:34,960 --> 01:03:38,360
I agree with you there's a

1771
01:03:38,360 --> 01:03:39,720
there's a lot of

1772
01:03:39,720 --> 01:03:42,520
bold, shall we say, marketing claims

1773
01:03:42,520 --> 01:03:44,720
Yeah. that

1774
01:03:44,720 --> 01:03:46,600
uh

1775
01:03:46,600 --> 01:03:48,720
funny in a way. I kind of see on one

1776
01:03:48,720 --> 01:03:51,040
hand it puts people off, but on the

1777
01:03:51,040 --> 01:03:52,760
other hand

1778
01:03:52,760 --> 01:03:54,920
I kind of see that some companies need

1779
01:03:54,920 --> 01:03:57,360
to do that almost to get the attention.

1780
01:03:57,360 --> 01:03:59,440
So there's a sort of uh

1781
01:03:59,440 --> 01:04:02,000
counterbalance um

1782
01:04:02,000 --> 01:04:04,800
and and I kind of wonder

1783
01:04:04,800 --> 01:04:06,840
well, firstly, whether

1784
01:04:06,840 --> 01:04:08,640
whether that's true. I mean, I think

1785
01:04:08,640 --> 01:04:10,800
it's crying out for a proper study that

1786
01:04:10,800 --> 01:04:13,160
does a more fairer comparison because I

1787
01:04:13,160 --> 01:04:14,680
know that just looking at the inference

1788
01:04:14,680 --> 01:04:15,800
time

1789
01:04:15,800 --> 01:04:17,000
alone

1790
01:04:17,000 --> 01:04:20,520
is is not the full picture. Right.

1791
01:04:20,520 --> 01:04:22,040
To the same point that I know that some

1792
01:04:22,040 --> 01:04:23,680
wind tunnel people hate it when you say,

1793
01:04:23,680 --> 01:04:25,160
"Oh, the cost of CFD is this and the

1794
01:04:25,160 --> 01:04:26,280
cost of a wind tunnel is this." And they

1795
01:04:26,280 --> 01:04:27,360
say, "Well, hold on, how long did you

1796
01:04:27,360 --> 01:04:30,160
take to mesh that?" Now, I know maybe

1797
01:04:30,160 --> 01:04:31,280
with your code you don't need to spend

1798
01:04:31,280 --> 01:04:33,080
as long meshing it, but there's always a

1799
01:04:33,080 --> 01:04:35,560
slight challenge of comparing things

1800
01:04:35,560 --> 01:04:38,160
like like. Um but what do you think

1801
01:04:38,160 --> 01:04:42,120
about the bolder claim around

1802
01:04:42,120 --> 01:04:44,400
sort of foundational

1803
01:04:44,400 --> 01:04:47,840
models? So the theory being that if I'm

1804
01:04:47,840 --> 01:04:50,080
an automotive customer and I always run

1805
01:04:50,080 --> 01:04:51,680
cars,

1806
01:04:51,680 --> 01:04:53,360
do you not perceive that with enough

1807
01:04:53,360 --> 01:04:55,840
high fidelity simulations

1808
01:04:55,840 --> 01:04:57,560
with enough sort of input-output

1809
01:04:57,560 --> 01:04:59,680
mappings that you could get to a point

1810
01:04:59,680 --> 01:05:02,040
where you could train a model that could

1811
01:05:02,040 --> 01:05:03,560
predict

1812
01:05:03,560 --> 01:05:05,000
a car?

1813
01:05:05,000 --> 01:05:07,160
You know, like is it a simply a data

1814
01:05:07,160 --> 01:05:09,840
problem? If there's enough data,

1815
01:05:09,840 --> 01:05:11,800
do you think you can ultimately

1816
01:05:11,800 --> 01:05:14,440
get to that point, or do you see it more

1817
01:05:14,440 --> 01:05:15,720
from a

1818
01:05:15,720 --> 01:05:16,680
sort of

1819
01:05:16,680 --> 01:05:18,800
physics enforcing

1820
01:05:18,800 --> 01:05:20,840
as in it's okay for low fidelity, but if

1821
01:05:20,840 --> 01:05:22,280
you really want to

1822
01:05:22,280 --> 01:05:24,000
get everything right, you still need the

1823
01:05:24,000 --> 01:05:25,200
high

1824
01:05:25,200 --> 01:05:26,440
fidelity?

1825
01:05:26,440 --> 01:05:28,920
Um

1826
01:05:29,040 --> 01:05:31,040
definitely Well,

1827
01:05:31,040 --> 01:05:33,200
that's a wonderful question, and let me

1828
01:05:33,200 --> 01:05:36,600
say a few things about it.

1829
01:05:36,600 --> 01:05:38,280
It's uh going to be a factor of the

1830
01:05:38,280 --> 01:05:40,760
level of accuracy you require. Sorry.

1831
01:05:40,760 --> 01:05:43,240
Don't worry.

1832
01:05:43,280 --> 01:05:44,640
Don't use the phone.

1833
01:05:44,640 --> 01:05:46,160
It doesn't do it for me. I don't know I

1834
01:05:46,160 --> 01:05:47,040
think

1835
01:05:47,040 --> 01:05:50,920
You could do it in the outtakes. Uh

1836
01:05:51,960 --> 01:05:55,720
I think the answer to that is is

1837
01:05:55,720 --> 01:05:58,600
multi-pronged. So,

1838
01:05:58,600 --> 01:06:02,120
let me uh let me start by saying that I

1839
01:06:02,120 --> 01:06:04,720
I do believe

1840
01:06:04,720 --> 01:06:07,560
that additional amount of data, vast

1841
01:06:07,560 --> 01:06:10,520
amount of data, are going to be helpful.

1842
01:06:10,520 --> 01:06:12,000
But, I think

1843
01:06:12,000 --> 01:06:15,600
whether you can replace a physics solver

1844
01:06:15,600 --> 01:06:20,000
by an ML AI sort of capability, it's

1845
01:06:20,000 --> 01:06:22,720
going to depend very strongly

1846
01:06:22,720 --> 01:06:24,600
on the level of accuracy that you

1847
01:06:24,600 --> 01:06:25,920
require.

1848
01:06:25,920 --> 01:06:27,560
I I'm a little biased. I come from the

1849
01:06:27,560 --> 01:06:29,720
aerospace industry where, you know, a

1850
01:06:29,720 --> 01:06:31,040
half a percent difference in the

1851
01:06:31,040 --> 01:06:32,880
prediction of drag and therefore fuel

1852
01:06:32,880 --> 01:06:36,080
burn of an aircraft is a huge number.

1853
01:06:36,080 --> 01:06:37,760
Mhm.

1854
01:06:37,760 --> 01:06:39,480
I have some hope

1855
01:06:39,480 --> 01:06:43,080
that if we had data for thousands or

1856
01:06:43,080 --> 01:06:44,800
hundreds of thousands or millions of

1857
01:06:44,800 --> 01:06:46,520
airplanes, we'll get there. But, I think

1858
01:06:46,520 --> 01:06:48,080
that's going to be a very expensive

1859
01:06:48,080 --> 01:06:49,640
proposition.

1860
01:06:49,640 --> 01:06:51,840
If you're designing a valve, you know,

1861
01:06:51,840 --> 01:06:54,680
for an irrigation system,

1862
01:06:54,680 --> 01:06:58,120
and you can tolerate five or 10% errors,

1863
01:06:58,120 --> 01:07:00,240
then I think there's going to be strong

1864
01:07:00,240 --> 01:07:03,320
potential use even without hundreds of

1865
01:07:03,320 --> 01:07:05,240
millions of data points, etc., etc.,

1866
01:07:05,240 --> 01:07:07,200
right? Uh Mhm. So, I do think the

1867
01:07:07,200 --> 01:07:09,680
accuracy level is going to be important.

1868
01:07:09,680 --> 01:07:10,360
Um

1869
01:07:10,360 --> 01:07:12,160
I do think the availability of massive

1870
01:07:12,160 --> 01:07:13,320
amounts of data is going to be

1871
01:07:13,320 --> 01:07:14,680
important.

1872
01:07:14,680 --> 01:07:17,120
In aerospace, I

1873
01:07:17,120 --> 01:07:18,480
mean at the end of the day, Neil, you

1874
01:07:18,480 --> 01:07:20,920
know this, right? The CFD solver is

1875
01:07:20,920 --> 01:07:23,360
solving a PDE

1876
01:07:23,360 --> 01:07:25,320
with a basis.

1877
01:07:25,320 --> 01:07:26,880
And because we have meshes that are very

1878
01:07:26,880 --> 01:07:28,120
fine,

1879
01:07:28,120 --> 01:07:30,400
these bases can represent all kinds of

1880
01:07:30,400 --> 01:07:32,760
features that are very small in size and

1881
01:07:32,760 --> 01:07:34,600
therefore, you know, there there's a

1882
01:07:34,600 --> 01:07:37,640
very good ability to represent the exact

1883
01:07:37,640 --> 01:07:39,440
solution of that PDE that you're

1884
01:07:39,440 --> 01:07:41,160
actually solving.

1885
01:07:41,160 --> 01:07:42,520
Here, what you're doing is you have a

1886
01:07:42,520 --> 01:07:44,560
non-linear basis that you're combining

1887
01:07:44,560 --> 01:07:47,400
in various different ways, but part of

1888
01:07:47,400 --> 01:07:51,200
the decrease in cost of the simulation

1889
01:07:51,200 --> 01:07:53,440
requires that the number of basis

1890
01:07:53,440 --> 01:07:56,200
vectors, let's say, is significantly

1891
01:07:56,200 --> 01:07:58,200
reduced. So, you're always going to be

1892
01:07:58,200 --> 01:07:59,560
trading this. You know, we've been doing

1893
01:07:59,560 --> 01:08:01,960
this from finite element bases to, you

1894
01:08:01,960 --> 01:08:03,680
know, wavelet transforms and various

1895
01:08:03,680 --> 01:08:05,600
other things. This is yet another

1896
01:08:05,600 --> 01:08:08,120
non-linear basis combination of various

1897
01:08:08,120 --> 01:08:10,200
different tools, which will have

1898
01:08:10,200 --> 01:08:12,680
potential use in many areas, but I don't

1899
01:08:12,680 --> 01:08:15,800
see it as a complete replacement of the

1900
01:08:15,800 --> 01:08:16,960
types of things that we're talking

1901
01:08:16,960 --> 01:08:18,200
about. I

1902
01:08:18,200 --> 01:08:21,240
Also, I have to tell you that the

1903
01:08:21,240 --> 01:08:23,600
I've lived this over the last 5 years.

1904
01:08:23,600 --> 01:08:25,000
When you go from executing one

1905
01:08:25,000 --> 01:08:29,799
calculation in 6 hours to 1 minute,

1906
01:08:29,799 --> 01:08:32,880
your your trade-off between speed of

1907
01:08:32,880 --> 01:08:35,359
execution of a surrogate model of some

1908
01:08:35,359 --> 01:08:39,600
kind and you know, the the actual

1909
01:08:39,600 --> 01:08:41,640
accuracy that you need changes, right?

1910
01:08:41,640 --> 01:08:44,799
So, and I I do think that that's going

1911
01:08:44,799 --> 01:08:47,040
to sort of be sort of coming after us

1912
01:08:47,040 --> 01:08:48,960
for a long, long time in some

1913
01:08:48,960 --> 01:08:50,600
applications where the accuracy is very,

1914
01:08:50,600 --> 01:08:53,000
very, very important.

1915
01:08:53,000 --> 01:08:54,839
In other applications, I think we'll

1916
01:08:54,839 --> 01:08:56,400
we'll transition to some of those models

1917
01:08:56,400 --> 01:08:58,440
more quickly. I think there's a dearth

1918
01:08:58,440 --> 01:09:00,520
of research

1919
01:09:00,520 --> 01:09:02,160
that is trying to understand

1920
01:09:02,160 --> 01:09:05,000
fundamentals of the physics

1921
01:09:05,000 --> 01:09:06,400
to then

1922
01:09:06,400 --> 01:09:08,680
introduce in a machine learning

1923
01:09:08,680 --> 01:09:11,400
methodology to extrapolate from the data

1924
01:09:11,400 --> 01:09:13,560
you learn from by using sort of the

1925
01:09:13,560 --> 01:09:15,000
commonality of the physics that was

1926
01:09:15,000 --> 01:09:17,640
actually learned. I'm excited about sort

1927
01:09:17,640 --> 01:09:20,240
of researching those areas that that

1928
01:09:20,240 --> 01:09:22,000
could have tremendous potential over the

1929
01:09:22,000 --> 01:09:23,600
next 10 years. So, there are a number of

1930
01:09:23,600 --> 01:09:25,000
people around the world trying these

1931
01:09:25,000 --> 01:09:26,120
ideas.

1932
01:09:26,120 --> 01:09:28,680
Uh I I almost see it as a

1933
01:09:28,680 --> 01:09:32,319
an ironic way that the very thing

1934
01:09:32,319 --> 01:09:34,160
that makes machine learning so

1935
01:09:34,160 --> 01:09:37,080
potentially transformative

1936
01:09:37,080 --> 01:09:38,400
um

1937
01:09:38,400 --> 01:09:40,319
which you could which some people may

1938
01:09:40,319 --> 01:09:42,960
think would challenge the need

1939
01:09:42,960 --> 01:09:45,440
for research into sort of high fidelity

1940
01:09:45,440 --> 01:09:50,040
or improved CFD is almost the motivation

1941
01:09:50,040 --> 01:09:51,720
for it. And the reason I say that is

1942
01:09:51,720 --> 01:09:53,000
because if you need so much training

1943
01:09:53,000 --> 01:09:54,200
data

1944
01:09:54,200 --> 01:09:55,920
the actual method you use to create the

1945
01:09:55,920 --> 01:09:59,560
training data becomes a huge cost. So,

1946
01:09:59,560 --> 01:10:01,480
if you can if your code can run five

1947
01:10:01,480 --> 01:10:02,560
times faster, means you could

1948
01:10:02,560 --> 01:10:04,160
potentially generate five times more

1949
01:10:04,160 --> 01:10:05,680
training examples and make your model

1950
01:10:05,680 --> 01:10:07,080
much better.

1951
01:10:07,080 --> 01:10:10,600
Now, whether the cost saving means that

1952
01:10:10,600 --> 01:10:11,760
why you're even doing that, you know,

1953
01:10:11,760 --> 01:10:12,920
you might as well just run the normal

1954
01:10:12,920 --> 01:10:14,320
simulation. That's I guess to be

1955
01:10:14,320 --> 01:10:15,920
determined.

1956
01:10:15,920 --> 01:10:19,320
Um it's it's it seems to be almost a

1957
01:10:19,320 --> 01:10:21,800
motivation for the higher fidelity fast

1958
01:10:21,800 --> 01:10:25,520
methods. I I'll make a

1959
01:10:25,520 --> 01:10:26,960
I'll make a prediction and then I'll

1960
01:10:26,960 --> 01:10:28,800
I'll give you an analogy.

1961
01:10:28,800 --> 01:10:29,720
The

1962
01:10:29,720 --> 01:10:31,760
The prediction is I think the the small

1963
01:10:31,760 --> 01:10:33,400
companies that are

1964
01:10:33,400 --> 01:10:35,320
simply

1965
01:10:35,320 --> 01:10:38,120
sort of taking simulation data from some

1966
01:10:38,120 --> 01:10:40,080
other provider

1967
01:10:40,080 --> 01:10:42,000
and doing

1968
01:10:42,000 --> 01:10:44,680
let's say ML AI any one of the

1969
01:10:44,680 --> 01:10:46,480
methodologies

1970
01:10:46,480 --> 01:10:48,400
um making the claims that we're talking

1971
01:10:48,400 --> 01:10:51,000
about are unlikely to be successful and

1972
01:10:51,000 --> 01:10:52,400
viable companies because there'll be a

1973
01:10:52,400 --> 01:10:53,640
number of use cases as we were

1974
01:10:53,640 --> 01:10:56,520
discussing before that could be useful,

1975
01:10:56,520 --> 01:10:57,960
but they're going to depend very heavily

1976
01:10:57,960 --> 01:11:00,920
on on on tight integration

1977
01:11:00,920 --> 01:11:02,880
with the simulation tools that are

1978
01:11:02,880 --> 01:11:05,400
producing the data in the first place.

1979
01:11:05,400 --> 01:11:07,680
And and they're not adding a significant

1980
01:11:07,680 --> 01:11:10,000
amount of IP in terms of coming up with

1981
01:11:10,000 --> 01:11:11,920
brand new methods that are not fully

1982
01:11:11,920 --> 01:11:14,240
published out there. And with PyTorch,

1983
01:11:14,240 --> 01:11:15,720
most people who know what they're doing

1984
01:11:15,720 --> 01:11:16,880
could actually code up in a couple

1985
01:11:16,880 --> 01:11:19,760
months, right? So So I I think there's

1986
01:11:19,760 --> 01:11:21,360
going to have to be much more

1987
01:11:21,360 --> 01:11:24,640
interaction between companies or tools

1988
01:11:24,640 --> 01:11:27,200
that do the AI and the machine learning

1989
01:11:27,200 --> 01:11:29,400
and simulation technology in order to be

1990
01:11:29,400 --> 01:11:31,520
able to sort of address many of these

1991
01:11:31,520 --> 01:11:33,920
different fields. So you you asked me a

1992
01:11:33,920 --> 01:11:36,360
question as to where I think this whole

1993
01:11:36,360 --> 01:11:38,240
thing is going. And again, I I don't

1994
01:11:38,240 --> 01:11:39,840
want to sound cynical at all. I think

1995
01:11:39,840 --> 01:11:41,560
I'm very upbeat about it for certain

1996
01:11:41,560 --> 01:11:43,720
uses. I

1997
01:11:43,720 --> 01:11:45,280
you know, as I get older, I can

1998
01:11:45,280 --> 01:11:46,880
pontificate about these kinds of things.

1999
01:11:46,880 --> 01:11:47,720
So

2000
01:11:47,720 --> 01:11:51,400
over my lifetime, I've seen NX, PVM and

2001
01:11:51,400 --> 01:11:54,240
MPI, you know, for for um

2002
01:11:54,240 --> 01:11:56,560
message passing sort of things. Then

2003
01:11:56,560 --> 01:11:59,120
then you could argue you you see single

2004
01:11:59,120 --> 01:12:01,840
core to multi-core processors to

2005
01:12:01,840 --> 01:12:03,720
Titanium processors to IBM cell

2006
01:12:03,720 --> 01:12:07,280
processors to GPU computing, etc. etc.

2007
01:12:07,280 --> 01:12:09,760
You never know which of those things are

2008
01:12:09,760 --> 01:12:12,280
really going to pan out and which ones

2009
01:12:12,280 --> 01:12:14,240
are going to just, you know, die and

2010
01:12:14,240 --> 01:12:16,160
wither on the vine. So

2011
01:12:16,160 --> 01:12:19,520
I think AI and ML is a broad uh topic

2012
01:12:19,520 --> 01:12:21,800
with many different elements. I think

2013
01:12:21,800 --> 01:12:23,600
some of those are going to be successful

2014
01:12:23,600 --> 01:12:25,280
and some of them are going to die and

2015
01:12:25,280 --> 01:12:27,120
wither on the vine because

2016
01:12:27,120 --> 01:12:28,440
uh you know, high-performance computing

2017
01:12:28,440 --> 01:12:30,520
simulations are going to be done. They

2018
01:12:30,520 --> 01:12:32,920
don't need millions of evaluations and

2019
01:12:32,920 --> 01:12:34,120
and therefore you're going to be better

2020
01:12:34,120 --> 01:12:35,640
off just doing it directly and not

2021
01:12:35,640 --> 01:12:37,080
having to worry about whether you have

2022
01:12:37,080 --> 01:12:39,440
to reevaluate, retrain, so on and so

2023
01:12:39,440 --> 01:12:41,040
forth. So

2024
01:12:41,040 --> 01:12:45,040
So yeah, it's it it's it's early uh

2025
01:12:45,040 --> 01:12:47,720
stages. And I think I think you can see,

2026
01:12:47,720 --> 01:12:49,640
as I told you, I'm convinced that

2027
01:12:49,640 --> 01:12:52,240
there's three, four, five uses where

2028
01:12:52,240 --> 01:12:54,400
this is going to be very, very powerful.

2029
01:12:54,400 --> 01:12:56,120
And I think there are others where it's

2030
01:12:56,120 --> 01:12:59,000
just hype and it'll go away.

2031
01:12:59,000 --> 01:13:00,960
One of the um

2032
01:13:00,960 --> 01:13:02,720
I think interesting things that you've

2033
01:13:02,720 --> 01:13:04,400
done and

2034
01:13:04,400 --> 01:13:05,560
and I guess there's a little bit of

2035
01:13:05,560 --> 01:13:08,000
analogy to the ML world, which is

2036
01:13:08,000 --> 01:13:09,760
this whole argument of open-source

2037
01:13:09,760 --> 01:13:11,680
versus closed source. What's the way to

2038
01:13:11,680 --> 01:13:13,840
advance things? Obviously, one of the

2039
01:13:13,840 --> 01:13:16,080
things I think you have made a large

2040
01:13:16,080 --> 01:13:18,680
contribution as well with your you know,

2041
01:13:18,680 --> 01:13:20,520
colleagues and people who who founded

2042
01:13:20,520 --> 01:13:24,440
the SU2 movement. Um

2043
01:13:24,440 --> 01:13:28,480
What do you think are the sort of

2044
01:13:28,840 --> 01:13:31,960
What can open-source codes achieve? What

2045
01:13:31,960 --> 01:13:34,120
can't they achieve? Well, you know, what

2046
01:13:34,120 --> 01:13:35,920
Where do you need to almost have a

2047
01:13:35,920 --> 01:13:37,680
commercial company

2048
01:13:37,680 --> 01:13:39,880
that an open-source couldn't do? Or I'm

2049
01:13:39,880 --> 01:13:41,520
I'm kind of always interested on what's

2050
01:13:41,520 --> 01:13:43,480
like the way to advance the science in a

2051
01:13:43,480 --> 01:13:45,280
way.

2052
01:13:45,280 --> 01:13:48,320
Yeah, so I came to open-source

2053
01:13:48,320 --> 01:13:50,880
reluctantly, I would say. We had always

2054
01:13:50,880 --> 01:13:53,080
given our codes that we had developed at

2055
01:13:53,080 --> 01:13:55,760
Stanford to anybody who wanted them, but

2056
01:13:55,760 --> 01:13:57,480
we said, "Here it is, not much

2057
01:13:57,480 --> 01:13:59,400
documentation, you're on your own,

2058
01:13:59,400 --> 01:14:00,720
right?"

2059
01:14:00,720 --> 01:14:01,920
So,

2060
01:14:01,920 --> 01:14:03,960
when I came back from NASA headquarters

2061
01:14:03,960 --> 01:14:07,280
in 2009, I had

2062
01:14:07,280 --> 01:14:10,320
a very talented sort of researcher in my

2063
01:14:10,320 --> 01:14:13,040
lab, Francisco Palacios, who uh you

2064
01:14:13,040 --> 01:14:14,480
probably have come across at some point

2065
01:14:14,480 --> 01:14:16,040
in your before then.

2066
01:14:16,040 --> 01:14:17,360
And

2067
01:14:17,360 --> 01:14:19,520
he convinced me that we had to ditch our

2068
01:14:19,520 --> 01:14:21,320
multi-block solvers and start doing

2069
01:14:21,320 --> 01:14:22,960
unstructured, and we said, "Okay, let's

2070
01:14:22,960 --> 01:14:25,160
do it." And then he and some of the

2071
01:14:25,160 --> 01:14:28,040
students were the ones who said,

2072
01:14:28,040 --> 01:14:30,360
"Let's put it on the open-source." And

2073
01:14:30,360 --> 01:14:31,400
I'll get back to your question in a

2074
01:14:31,400 --> 01:14:34,200
moment, but just for context.

2075
01:14:34,200 --> 01:14:35,840
And at the time I was very concerned

2076
01:14:35,840 --> 01:14:37,760
about it. I was like, "Well, it carries

2077
01:14:37,760 --> 01:14:39,800
the Stanford name, so we cannot have

2078
01:14:39,800 --> 01:14:41,480
something that's crappy."

2079
01:14:41,480 --> 01:14:43,160
Um

2080
01:14:43,160 --> 01:14:44,800
it's something that people are going to

2081
01:14:44,800 --> 01:14:46,960
be to a lot of questions, so we have to

2082
01:14:46,960 --> 01:14:48,600
have documentation and various other

2083
01:14:48,600 --> 01:14:51,760
things. You know, it's not clear that in

2084
01:14:51,760 --> 01:14:53,360
a university environment with people who

2085
01:14:53,360 --> 01:14:55,800
come and go over time, we're going to be

2086
01:14:55,800 --> 01:14:57,480
able to sustain it over a long period of

2087
01:14:57,480 --> 01:14:58,760
time. So, it was important to start

2088
01:14:58,760 --> 01:15:00,000
enlisting other colleagues,

2089
01:15:00,000 --> 01:15:02,000
collaborators around the world.

2090
01:15:02,000 --> 01:15:04,240
Um but I had learned a very valuable

2091
01:15:04,240 --> 01:15:06,160
management lesson at NASA from somebody

2092
01:15:06,160 --> 01:15:07,840
who was my senior technical advisor

2093
01:15:07,840 --> 01:15:08,720
there.

2094
01:15:08,720 --> 01:15:10,080
And that is that you never say no

2095
01:15:10,080 --> 01:15:13,560
because you say yes if. And we organized

2096
01:15:13,560 --> 01:15:14,960
ourselves

2097
01:15:14,960 --> 01:15:16,600
you know, to make it happen and we put

2098
01:15:16,600 --> 01:15:18,360
it out in the open-source and we wish

2099
01:15:18,360 --> 01:15:19,600
for the best.

2100
01:15:19,600 --> 01:15:21,680
And at the time to be honest, my

2101
01:15:21,680 --> 01:15:24,320
ambition for SU2 was to make sure my

2102
01:15:24,320 --> 01:15:26,280
students at Stanford didn't have to

2103
01:15:26,280 --> 01:15:29,160
redevelop technology every time that was

2104
01:15:29,160 --> 01:15:32,080
not intrinsic or or exciting or

2105
01:15:32,080 --> 01:15:34,560
interesting for their own research. That

2106
01:15:34,560 --> 01:15:36,120
was the motivation.

2107
01:15:36,120 --> 01:15:38,880
Um soon after we released it, several

2108
01:15:38,880 --> 01:15:40,840
key universities around the world jumped

2109
01:15:40,840 --> 01:15:43,760
in and I I I have to say the value of

2110
01:15:43,760 --> 01:15:45,360
open-source

2111
01:15:45,360 --> 01:15:48,200
was twofold. One One was about building

2112
01:15:48,200 --> 01:15:51,040
community of people who had different

2113
01:15:51,040 --> 01:15:53,280
areas of expertise but were like-minded

2114
01:15:53,280 --> 01:15:54,960
about developing new capabilities and

2115
01:15:54,960 --> 01:15:56,400
sharing it.

2116
01:15:56,400 --> 01:15:59,240
Um and I would say it was also very

2117
01:15:59,240 --> 01:16:00,920
important uh

2118
01:16:00,920 --> 01:16:03,480
to make sure that we generated an a new

2119
01:16:03,480 --> 01:16:05,360
generation of graduate students who were

2120
01:16:05,360 --> 01:16:08,480
trained in these five types of tools who

2121
01:16:08,480 --> 01:16:10,360
then go out to industry and sort of do

2122
01:16:10,360 --> 01:16:13,640
amazing things. So, so to me SU2 was the

2123
01:16:13,640 --> 01:16:15,320
most wonderful experience that was

2124
01:16:15,320 --> 01:16:16,840
completely serendipitous and it

2125
01:16:16,840 --> 01:16:18,480
continues to this day.

2126
01:16:18,480 --> 01:16:20,360
Mhm. But

2127
01:16:20,360 --> 01:16:22,400
open-source at least at the level of

2128
01:16:22,400 --> 01:16:24,600
SU2, you know more about OpenFOAM than I

2129
01:16:24,600 --> 01:16:26,200
do although we we've interacted with the

2130
01:16:26,200 --> 01:16:28,200
OpenFOAM team many times.

2131
01:16:28,200 --> 01:16:30,640
Um

2132
01:16:30,720 --> 01:16:32,320
if you think the most important thing is

2133
01:16:32,320 --> 01:16:33,840
X

2134
01:16:33,840 --> 01:16:36,440
and the universities have funding for Y

2135
01:16:36,440 --> 01:16:38,200
Y gets done

2136
01:16:38,200 --> 01:16:39,800
and X, you know, maybe we make a little

2137
01:16:39,800 --> 01:16:41,880
bit of progress if we organize ourselves

2138
01:16:41,880 --> 01:16:43,400
and we put a foundation together and

2139
01:16:43,400 --> 01:16:44,560
various things, we tried all those

2140
01:16:44,560 --> 01:16:45,560
things.

2141
01:16:45,560 --> 01:16:46,960
So, what ends up happening is that the

2142
01:16:46,960 --> 01:16:49,160
progress is slower.

2143
01:16:49,160 --> 01:16:52,800
You know, what you can do in industry,

2144
01:16:52,800 --> 01:16:54,960
with venture capital funds, with focus,

2145
01:16:54,960 --> 01:16:57,440
with ex- experienced people who are not

2146
01:16:57,440 --> 01:16:59,040
coding, you know, the solver for the

2147
01:16:59,040 --> 01:17:01,480
first time, etc., etc., means that you

2148
01:17:01,480 --> 01:17:03,280
can do things much more quickly, more

2149
01:17:03,280 --> 01:17:05,800
professionally, more robustly, uh and

2150
01:17:05,800 --> 01:17:07,440
you can do end-to-end solutions. It's

2151
01:17:07,440 --> 01:17:09,160
very hard to do that in open-source, as

2152
01:17:09,160 --> 01:17:11,440
you know. Um

2153
01:17:11,440 --> 01:17:13,000
at the same time,

2154
01:17:13,000 --> 01:17:14,640
it's how you train people to be able to

2155
01:17:14,640 --> 01:17:16,160
do great things, because they have

2156
01:17:16,160 --> 01:17:18,320
access to every line of code of a

2157
01:17:18,320 --> 01:17:20,560
state-of-the-art algorithm, a

2158
01:17:20,560 --> 01:17:22,720
state-of-the-art physics model, a

2159
01:17:22,720 --> 01:17:24,640
state-of-the-art multiphysics coupling,

2160
01:17:24,640 --> 01:17:26,080
a state-of-the-art automatic

2161
01:17:26,080 --> 01:17:29,120
differentiation tool. So, I think it in

2162
01:17:29,120 --> 01:17:31,840
in academia, we have to figure out how

2163
01:17:31,840 --> 01:17:33,480
to leverage open-source tools to train

2164
01:17:33,480 --> 01:17:35,640
our students to be the next generation

2165
01:17:35,640 --> 01:17:37,040
of people who do amazing things like

2166
01:17:37,040 --> 01:17:38,680
we're doing at Luminary.

2167
01:17:38,680 --> 01:17:40,800
Mhm. And without those people that

2168
01:17:40,800 --> 01:17:42,400
understand the details, you don't get to

2169
01:17:42,400 --> 01:17:44,760
do amazing things. So, so there there's

2170
01:17:44,760 --> 01:17:46,800
a tremendous value of open-source. At

2171
01:17:46,800 --> 01:17:48,640
the same time, you know, I'm back in

2172
01:17:48,640 --> 01:17:50,560
academia now.

2173
01:17:50,560 --> 01:17:51,880
And

2174
01:17:51,880 --> 01:17:54,880
I know academia cannot compete

2175
01:17:54,880 --> 01:17:56,360
with the level of talent we put together

2176
01:17:56,360 --> 01:17:57,760
at a company like Luminary, and I

2177
01:17:57,760 --> 01:17:59,760
imagine other companies that may be

2178
01:17:59,760 --> 01:18:02,000
trying to do similar things. And

2179
01:18:02,000 --> 01:18:04,440
um it is a good question that I don't

2180
01:18:04,440 --> 01:18:06,240
have a good answer to.

2181
01:18:06,240 --> 01:18:09,040
That's what academia should be doing to

2182
01:18:09,040 --> 01:18:11,880
add value on how could companies team up

2183
01:18:11,880 --> 01:18:14,160
with academic teams

2184
01:18:14,160 --> 01:18:16,600
to allow the use of these established

2185
01:18:16,600 --> 01:18:18,920
tools that that are, you know,

2186
01:18:18,920 --> 01:18:21,200
non-research value,

2187
01:18:21,200 --> 01:18:23,720
but enable the research in order for

2188
01:18:23,720 --> 01:18:25,600
academia to continue to be relevant. And

2189
01:18:25,600 --> 01:18:27,280
that's something that I'm I'm struggling

2190
01:18:27,280 --> 01:18:29,040
with a little bit right now, and need to

2191
01:18:29,040 --> 01:18:31,360
continue sort of pushing on. I

2192
01:18:31,360 --> 01:18:32,840
Continuing open-source is one of the

2193
01:18:32,840 --> 01:18:34,600
options, right? And then and that I'm

2194
01:18:34,600 --> 01:18:37,360
very committed to the future of SU2.

2195
01:18:37,360 --> 01:18:40,240
but but there may be other models that

2196
01:18:40,240 --> 01:18:42,800
allow us to continue to push the

2197
01:18:42,800 --> 01:18:45,240
research boundaries in academia with the

2198
01:18:45,240 --> 01:18:47,800
aid of commercial tools. And how how do

2199
01:18:47,800 --> 01:18:49,040
we do that? I think it's a good

2200
01:18:49,040 --> 01:18:50,800
question. I don't know if you have any

2201
01:18:50,800 --> 01:18:51,560
any thoughts

2202
01:18:51,560 --> 01:18:53,800
I don't know about that. Yeah, it's it's

2203
01:18:53,800 --> 01:18:54,760
um

2204
01:18:54,760 --> 01:18:56,480
it's also the case

2205
01:18:56,480 --> 01:18:59,480
I mean there's always a healthy um

2206
01:18:59,480 --> 01:19:01,520
movement, I guess, cuz you as you say

2207
01:19:01,520 --> 01:19:03,080
you need

2208
01:19:03,080 --> 01:19:05,240
if all of academia uses commercial tools

2209
01:19:05,240 --> 01:19:06,720
with no access to the source code,

2210
01:19:06,720 --> 01:19:09,000
they're never going to be able to learn

2211
01:19:09,000 --> 01:19:11,720
how to do the coding that is to develop

2212
01:19:11,720 --> 01:19:15,240
it. Right. Um but if the code if I think

2213
01:19:15,240 --> 01:19:16,960
you said it very well,

2214
01:19:16,960 --> 01:19:18,640
if every student has to start from the

2215
01:19:18,640 --> 01:19:19,920
beginning,

2216
01:19:19,920 --> 01:19:22,520
they never really get to focus on the

2217
01:19:22,520 --> 01:19:24,720
novelty bit. So, there needs to be a

2218
01:19:24,720 --> 01:19:27,960
foundation and so that it naturally

2219
01:19:27,960 --> 01:19:30,800
motivates the desire for codes, you

2220
01:19:30,800 --> 01:19:33,160
know, to become more

2221
01:19:33,160 --> 01:19:35,440
as a platform, I guess.

2222
01:19:35,440 --> 01:19:37,400
It's what you build then on top of it.

2223
01:19:37,400 --> 01:19:39,400
And that's why I guess the PyTorch sort

2224
01:19:39,400 --> 01:19:40,440
of thing is interesting that there's a

2225
01:19:40,440 --> 01:19:43,120
framework that CFD doesn't have, does

2226
01:19:43,120 --> 01:19:44,360
it? I mean, if you look at ML, most

2227
01:19:44,360 --> 01:19:45,920
people use PyTorch or TensorFlow and

2228
01:19:45,920 --> 01:19:47,440
build on top of it.

2229
01:19:47,440 --> 01:19:48,640
We

2230
01:19:48,640 --> 01:19:49,920
well, I I guess you could argue there's

2231
01:19:49,920 --> 01:19:51,600
some linear solvers and things like that

2232
01:19:51,600 --> 01:19:53,600
that people try and build on, but we we

2233
01:19:53,600 --> 01:19:55,320
haven't been as

2234
01:19:55,320 --> 01:19:57,840
coordinated, maybe. Yeah. Um Our

2235
01:19:57,840 --> 01:20:01,240
abstraction layer

2236
01:20:01,240 --> 01:20:04,240
has not been created as effectively as

2237
01:20:04,240 --> 01:20:07,720
AI/ML has done theirs. But of course, it

2238
01:20:07,720 --> 01:20:09,600
their theirs is a much much bigger

2239
01:20:09,600 --> 01:20:12,720
market, right? So, so there is there is

2240
01:20:12,720 --> 01:20:16,240
a very very high value to creating those

2241
01:20:16,240 --> 01:20:18,360
abstraction layers for execution in any

2242
01:20:18,360 --> 01:20:21,040
platform um than there is for CFD,

2243
01:20:21,040 --> 01:20:22,640
right? Not many people around the world

2244
01:20:22,640 --> 01:20:24,680
write CFD solvers.

2245
01:20:24,680 --> 01:20:27,680
So, um maybe a final question. So, you

2246
01:20:27,680 --> 01:20:30,480
you were the author of the 2030 report.

2247
01:20:30,480 --> 01:20:32,960
So, if you were going to now

2248
01:20:32,960 --> 01:20:35,680
put down what you think is this s-

2249
01:20:35,680 --> 01:20:37,280
2050

2250
01:20:37,280 --> 01:20:38,760
Yeah. What would you put as the grand

2251
01:20:38,760 --> 01:20:41,320
challenges? The stuff that can't be done

2252
01:20:41,320 --> 01:20:42,480
now

2253
01:20:42,480 --> 01:20:46,520
and are the things that still we need to

2254
01:20:46,520 --> 01:20:48,400
put as our grand challenges? If we

2255
01:20:48,400 --> 01:20:50,640
assume that the goals of 2030 are

2256
01:20:50,640 --> 01:20:52,920
achieved in 2030 or before, right? Let's

2257
01:20:52,920 --> 01:20:54,880
assume, like I mentioned, that inner

2258
01:20:54,880 --> 01:20:59,400
flow, inner workflow Mhm. is is fast, is

2259
01:20:59,400 --> 01:21:01,960
accurate, it always works, you know,

2260
01:21:01,960 --> 01:21:03,840
it's scalable, it can be executed

2261
01:21:03,840 --> 01:21:05,960
anywhere you want. So, let's say we've

2262
01:21:05,960 --> 01:21:07,920
forgotten about that.

2263
01:21:07,920 --> 01:21:10,040
I think the challenge, the the next

2264
01:21:10,040 --> 01:21:12,760
challenge is the outer flows

2265
01:21:12,760 --> 01:21:15,320
and how we're going to achieve them.

2266
01:21:15,320 --> 01:21:16,920
So, the outer flows for me are, you

2267
01:21:16,920 --> 01:21:18,480
know, always optimization, design

2268
01:21:18,480 --> 01:21:19,760
optimization, uncertainty

2269
01:21:19,760 --> 01:21:21,040
quantification, design and their

2270
01:21:21,040 --> 01:21:23,520
uncertainty, AI/ML, you know, those

2271
01:21:23,520 --> 01:21:26,720
things that require repeated evaluation

2272
01:21:26,720 --> 01:21:28,720
of the actual, you know, physical

2273
01:21:28,720 --> 01:21:30,280
models, right? So

2274
01:21:30,280 --> 01:21:32,680
So, I think it's all about that. I think

2275
01:21:32,680 --> 01:21:34,720
it's about multiphysics.

2276
01:21:34,720 --> 01:21:37,040
It's about judicious use of ML/AI,

2277
01:21:37,040 --> 01:21:39,440
right? The um

2278
01:21:39,440 --> 01:21:41,000
I think those are the challenges, but

2279
01:21:41,000 --> 01:21:43,160
then there are many, many challenges of

2280
01:21:43,160 --> 01:21:44,880
how we get there.

2281
01:21:44,880 --> 01:21:48,400
Right? Then Mhm. um

2282
01:21:48,400 --> 01:21:49,880
I strongly believe we're going to have

2283
01:21:49,880 --> 01:21:54,480
GPUs around for another 6 to 10 years.

2284
01:21:54,480 --> 01:21:55,560
That's the interesting one, isn't it?

2285
01:21:55,560 --> 01:21:58,560
Hardware, because it feels like

2286
01:21:58,560 --> 01:22:01,240
it's been machine learning that has made

2287
01:22:01,240 --> 01:22:03,080
GPUs. It's nothing to do with CFD, is

2288
01:22:03,080 --> 01:22:04,480
it, really? They didn't We haven't

2289
01:22:04,480 --> 01:22:06,040
suddenly got more GPUs cuz someone got,

2290
01:22:06,040 --> 01:22:07,680
"Oh, I really like fluid dynamics and

2291
01:22:07,680 --> 01:22:09,680
they're helping wind tunnels." It's

2292
01:22:09,680 --> 01:22:10,360
It's just come back

2293
01:22:10,360 --> 01:22:12,360
learning, it was video gaming. It's not

2294
01:22:12,360 --> 01:22:13,880
It wasn't the machine learning and it

2295
01:22:13,880 --> 01:22:15,680
wasn't computing, either, right? So

2296
01:22:15,680 --> 01:22:17,680
Yeah. um

2297
01:22:17,680 --> 01:22:19,400
We're going to run out of

2298
01:22:19,400 --> 01:22:21,280
ideas and and how to make these things

2299
01:22:21,280 --> 01:22:23,840
faster at some point with GPUs, as well,

2300
01:22:23,840 --> 01:22:24,880
right? The

2301
01:22:24,880 --> 01:22:26,520
Uh we can still do a few things. So,

2302
01:22:26,520 --> 01:22:28,000
there's there's a lot of technology that

2303
01:22:28,000 --> 01:22:30,480
I'm sure you know about um that gives me

2304
01:22:30,480 --> 01:22:33,000
confidence that that it's a stable model

2305
01:22:33,000 --> 01:22:34,600
for at least a good number of years,

2306
01:22:34,600 --> 01:22:35,920
right? Um

2307
01:22:35,920 --> 01:22:38,200
Um but there's going to be other ways in

2308
01:22:38,200 --> 01:22:40,480
which we use artificial intelligence. Uh

2309
01:22:40,480 --> 01:22:41,960
there's going to be quantum computing

2310
01:22:41,960 --> 01:22:43,160
coming along. I mean, if we're talking

2311
01:22:43,160 --> 01:22:45,880
about another 20 years, right? 2030 CFD

2312
01:22:45,880 --> 01:22:48,520
was written in essentially 2012. So, it

2313
01:22:48,520 --> 01:22:50,600
was almost a 20-year prediction.

2314
01:22:50,600 --> 01:22:52,640
So, we're talking about 20 more years.

2315
01:22:52,640 --> 01:22:54,360
It's not out of the scope that many of

2316
01:22:54,360 --> 01:22:56,640
these things will But would you not say

2317
01:22:56,640 --> 01:22:57,440
that

2318
01:22:57,440 --> 01:22:59,080
the irony

2319
01:22:59,080 --> 01:23:00,440
which goes back to right at the

2320
01:23:00,440 --> 01:23:02,080
beginning when I said about what's the

2321
01:23:02,080 --> 01:23:04,680
difference between design then and now

2322
01:23:04,680 --> 01:23:07,960
is it takes people 20 years to change

2323
01:23:07,960 --> 01:23:10,640
the way they work sometimes. That like

2324
01:23:10,640 --> 01:23:13,080
you know, the technology arguably

2325
01:23:13,080 --> 01:23:15,320
that you as a company are building in as

2326
01:23:15,320 --> 01:23:17,640
you said you're doing it now because it

2327
01:23:17,640 --> 01:23:19,520
has reached maturity

2328
01:23:19,520 --> 01:23:21,840
and then maybe over the next 10 years

2329
01:23:21,840 --> 01:23:25,800
people will start to use it. Yeah. But

2330
01:23:25,800 --> 01:23:28,440
it takes time. So, maybe the 2050 is

2331
01:23:28,440 --> 01:23:30,440
just that everybody will be using the

2332
01:23:30,440 --> 01:23:32,600
stuff available now

2333
01:23:32,600 --> 01:23:35,000
in production. I don't know, you know.

2334
01:23:35,000 --> 01:23:37,960
There's a human time constant that is

2335
01:23:37,960 --> 01:23:40,280
very important because it takes time to

2336
01:23:40,280 --> 01:23:42,840
build experience, build processes, so on

2337
01:23:42,840 --> 01:23:44,000
and so forth. Maybe there'll be

2338
01:23:44,000 --> 01:23:46,400
technology that accelerate that, but

2339
01:23:46,400 --> 01:23:49,040
that's been a constant over time.

2340
01:23:49,040 --> 01:23:50,560
Mhm. But there's also something that I

2341
01:23:50,560 --> 01:23:52,920
would say about the human beings is we

2342
01:23:52,920 --> 01:23:55,040
never stop innovating.

2343
01:23:55,040 --> 01:23:57,560
So, I I think we'll be using So, the the

2344
01:23:57,560 --> 01:23:59,640
state of the art in 2050 will be that

2345
01:23:59,640 --> 01:24:02,120
nobody will be thinking about

2346
01:24:02,120 --> 01:24:03,640
just doing a single simulation. They'll

2347
01:24:03,640 --> 01:24:05,040
be running a thousand simulations

2348
01:24:05,040 --> 01:24:07,640
simultaneously in 2 minutes. They'll be

2349
01:24:07,640 --> 01:24:09,240
real models about it and they'll be

2350
01:24:09,240 --> 01:24:11,240
querying them and you know, our vision

2351
01:24:11,240 --> 01:24:14,480
at Luminary was always the the Jarvis uh

2352
01:24:14,480 --> 01:24:15,960
Iron Man vision. I don't know if you're

2353
01:24:15,960 --> 01:24:18,200
familiar with that. It's this little AI

2354
01:24:18,200 --> 01:24:19,960
that Tony Stark has where you're like,

2355
01:24:19,960 --> 01:24:23,280
"Oh, I want to design a new rocket.

2356
01:24:23,280 --> 01:24:26,200
Yeah. here's three options you have. So

2357
01:24:26,200 --> 01:24:29,000
So, my guess is that people in 2050 will

2358
01:24:29,000 --> 01:24:31,120
be doing ensembles of simulations and

2359
01:24:31,120 --> 01:24:33,360
extracting information from those in

2360
01:24:33,360 --> 01:24:35,800
both design, uncertainty, risks, so on

2361
01:24:35,800 --> 01:24:37,280
and so forth, and that will become

2362
01:24:37,280 --> 01:24:39,280
commonplace. These are things that

2363
01:24:39,280 --> 01:24:41,080
academia has been sort of working on for

2364
01:24:41,080 --> 01:24:43,000
the last 10 years at least, I would say,

2365
01:24:43,000 --> 01:24:44,880
and and they're they're going to be

2366
01:24:44,880 --> 01:24:46,440
transitioned to become commonplace, but

2367
01:24:46,440 --> 01:24:48,400
there'll be something else beyond that.

2368
01:24:48,400 --> 01:24:49,640
Right? So,

2369
01:24:49,640 --> 01:24:50,880
I hope so.

2370
01:24:50,880 --> 01:24:52,560
It's something to work with.

2371
01:24:52,560 --> 01:24:53,480
Yeah.

2372
01:24:53,480 --> 01:24:54,480
That's it.

2373
01:24:54,480 --> 01:24:57,560
But but I I I have tremendous faith that

2374
01:24:57,560 --> 01:25:00,000
we'll keep innovating and however slowly

2375
01:25:00,000 --> 01:25:01,920
for those of us who are involved in this

2376
01:25:01,920 --> 01:25:03,760
business, you know, we see it day-to-day

2377
01:25:03,760 --> 01:25:05,120
and it seems slow, but when you look

2378
01:25:05,120 --> 01:25:06,840
back 10 20 years, you're like, "Wow, we

2379
01:25:06,840 --> 01:25:09,080
actually got some stuff done, right?"

2380
01:25:09,080 --> 01:25:10,960
There's been some step changes along the

2381
01:25:10,960 --> 01:25:12,720
way. There'll continue to be step

2382
01:25:12,720 --> 01:25:15,080
changes along the way by the year 2050.

2383
01:25:15,080 --> 01:25:17,320
So, I'm I'm quite bullish on the uh

2384
01:25:17,320 --> 01:25:19,320
on the future of computational driven

2385
01:25:19,320 --> 01:25:21,200
science. I think it's largely going to

2386
01:25:21,200 --> 01:25:24,400
replace physical experimentation by 2050

2387
01:25:24,400 --> 01:25:26,000
for sure in many engineering

2388
01:25:26,000 --> 01:25:29,080
disciplines, let's say. So, Yeah. Yeah.

2389
01:25:29,080 --> 01:25:31,000
Well, thank you for doing this. Uh it's

2390
01:25:31,000 --> 01:25:33,160
been lovely to chat to you, and I really

2391
01:25:33,160 --> 01:25:35,840
do I am so pleased to see, you know, the

2392
01:25:35,840 --> 01:25:37,480
Luminary Cloud that it's been launched

2393
01:25:37,480 --> 01:25:39,800
now, and I must have been a fantastic

2394
01:25:39,800 --> 01:25:42,240
excitement, stress journey, but I really

2395
01:25:42,240 --> 01:25:44,200
do wish you all the best and and hope it

2396
01:25:44,200 --> 01:25:47,160
is a a big success. People check it out

2397
01:25:47,160 --> 01:25:48,640
and um

2398
01:25:48,640 --> 01:25:51,120
hope we get to meet in person again

2399
01:25:51,120 --> 01:25:54,080
point at some conference and uh have a

2400
01:25:54,080 --> 01:25:55,840
drink or a coffee.

2401
01:25:55,840 --> 01:25:58,520
I would love to do that. Um thank you

2402
01:25:58,520 --> 01:26:01,000
for doing the podcast and inviting me to

2403
01:26:01,000 --> 01:26:03,200
the podcast. Thanks for the kind words

2404
01:26:03,200 --> 01:26:06,080
about Luminary. We really do think that

2405
01:26:06,080 --> 01:26:07,760
there's a nugget there that can help

2406
01:26:07,760 --> 01:26:10,800
change the way we do engineering, and

2407
01:26:10,800 --> 01:26:13,440
and it's always going to be fed by new

2408
01:26:13,440 --> 01:26:15,320
ideas that are going to come out from

2409
01:26:15,320 --> 01:26:17,880
academia. So, our academic colleagues

2410
01:26:17,880 --> 01:26:20,840
are are also, you know, amazing people

2411
01:26:20,840 --> 01:26:22,520
who have made this this current

2412
01:26:22,520 --> 01:26:24,240
state-of-the-art

2413
01:26:24,240 --> 01:26:25,720
sort of happen and

2414
01:26:25,720 --> 01:26:27,800
this is this going to continue along and

2415
01:26:27,800 --> 01:26:30,400
I think you and I will have many coffees

2416
01:26:30,400 --> 01:26:32,400
and good opportunities to think about

2417
01:26:32,400 --> 01:26:34,720
what the future might bring. Yeah,

2418
01:26:34,720 --> 01:26:37,160
awesome. Thanks very much. Neil, it's a

2419
01:26:37,160 --> 01:26:38,480
real pleasure. Thank you for the

2420
01:26:38,480 --> 01:26:40,200
invitation and, you know, have a nice

2421
01:26:40,200 --> 01:26:42,840
evening there.

2422
01:27:05,800 --> 01:27:07,920
Yeah.