1
00:00:00,240 --> 00:00:02,360
Hi, and welcome to the Neil
Ashton Podcast.

2
00:00:03,000 --> 00:00:05,880
In each episode, we explained
some of the fascinating ways

3
00:00:05,880 --> 00:00:09,080
that science and engineering are
changing the world around us.

4
00:00:09,760 --> 00:00:12,720
We talked to leading engineers
from elite level sports like

5
00:00:12,800 --> 00:00:16,840
cycling and Formula One to some
of the world's top academics to

6
00:00:16,840 --> 00:00:20,960
understand how fluid dynamics,
machine learning, supercomputing

7
00:00:21,360 --> 00:00:22,960
are bringing in a new era of
discovery.

8
00:00:23,920 --> 00:00:27,040
We also hear some of their life
stories, their career advice,

9
00:00:27,640 --> 00:00:30,040
the lessons they've learned on
the way that I hope will be

10
00:00:30,040 --> 00:00:33,800
helpful to you too.
So sit back and enjoy this

11
00:00:33,800 --> 00:00:40,840
episode.
Hi, and welcome back to the Neil

12
00:00:40,840 --> 00:00:44,040
Ashton Podcast.
So today's guest is Professor

13
00:00:44,040 --> 00:00:48,040
Ricardo Vinu Isa.
Ricardo is currently a Professor

14
00:00:48,040 --> 00:00:51,920
of Aerospace Engineering at the
University of Michigan, where he

15
00:00:51,920 --> 00:00:54,120
leads research at the
intersection of machine

16
00:00:54,120 --> 00:00:57,440
learning, fluid mechanics,
turbulence, and artificial

17
00:00:57,440 --> 00:01:00,480
intelligence.
Before joining Michigan, he was

18
00:01:00,480 --> 00:01:03,720
a professor at KTH Royal
Institute of Technology in

19
00:01:03,720 --> 00:01:06,400
Stockholm, one of Europe's
leading engineering

20
00:01:06,400 --> 00:01:10,240
universities, and he built a
internationally recognized

21
00:01:10,240 --> 00:01:14,400
research program focused on
turbulent flow control and

22
00:01:14,400 --> 00:01:17,360
data-driven methods for fluid
mechanics.

23
00:01:18,600 --> 00:01:21,040
He's definitely established
himself as one of the leading

24
00:01:21,040 --> 00:01:24,640
researchers looking into this.
And one of the topics that we

25
00:01:24,640 --> 00:01:27,800
bring on today, of course, is
around like explainable AI,

26
00:01:27,800 --> 00:01:30,600
causality, reinforcement
learning, reduced order

27
00:01:30,600 --> 00:01:35,440
modelling, and of course the
topic of the moment foundation

28
00:01:35,440 --> 00:01:40,240
models for fluid mechanics.
He's a long people like Steve

29
00:01:40,240 --> 00:01:42,080
Brunton.
He's done a lot to really

30
00:01:42,080 --> 00:01:46,600
educate the community, also the
public in terms of these fluid

31
00:01:46,600 --> 00:01:48,760
mechanics and how machine
learning can be used for.

32
00:01:48,760 --> 00:01:52,240
It's really, he's got some great
YouTube videos and explanations.

33
00:01:52,680 --> 00:01:55,720
And what I particularly enjoy
about his work is he doesn't

34
00:01:55,720 --> 00:01:58,800
just ask whether AI can make
predictions faster.

35
00:01:59,160 --> 00:02:02,920
He's looking at the question,
can AI help us to understand the

36
00:02:02,920 --> 00:02:06,920
mechanics of turbulence and to
ultimate accelerate scientific

37
00:02:06,920 --> 00:02:09,360
discovery itself.
So, you know, in this

38
00:02:09,360 --> 00:02:12,320
discussion, which as with any of
them, is never long enough to

39
00:02:12,320 --> 00:02:15,520
fully cover everything, you
know, we dived into some of the

40
00:02:15,520 --> 00:02:17,760
bigger questions facing CFT
today.

41
00:02:17,760 --> 00:02:21,640
So you know whether fluid
mechanics will have its ChatGPT

42
00:02:21,640 --> 00:02:24,720
moment, how close are we to
foundation models can be

43
00:02:24,720 --> 00:02:28,280
generalized across different
flow problems, and why he's

44
00:02:28,280 --> 00:02:31,880
actually more optimistic about
that possibility than than ever

45
00:02:31,880 --> 00:02:35,040
before.
As I mentioned before, the

46
00:02:35,040 --> 00:02:39,400
explainable AI causality,
understanding where they come

47
00:02:39,400 --> 00:02:42,000
from.
Can you explain how AI is

48
00:02:42,000 --> 00:02:44,920
getting to it?
And one of the things that we

49
00:02:44,920 --> 00:02:48,520
also look about is focusing on
the role of reinforcement

50
00:02:48,520 --> 00:02:52,720
learning in terms of
optimization and control and not

51
00:02:52,720 --> 00:02:55,320
just using AI as a pure
predictor.

52
00:02:55,440 --> 00:02:58,400
And then towards the end, we, we
sort of zoom out and talk about

53
00:02:58,400 --> 00:03:01,280
the future of the field, agentic
AI, autonomous scientific

54
00:03:01,280 --> 00:03:04,960
discovery, and also how
universities should be adapting

55
00:03:04,960 --> 00:03:09,280
education given that AI is
changing the whole landscape.

56
00:03:09,280 --> 00:03:12,840
So I, I think this is one of
these conversations that, you

57
00:03:12,840 --> 00:03:17,040
know, I learned a lot from it
and, and I hope you do too.

58
00:03:17,040 --> 00:03:20,920
So sit back and enjoy this
episode with Professor Vinu Isa.

59
00:03:21,280 --> 00:03:23,080
Thanks very much for agreeing to
do this.

60
00:03:23,640 --> 00:03:28,160
You are definitely on the list.
Of people who everyone says, oh,

61
00:03:28,160 --> 00:03:30,800
you should be speaking to him.
You know, I read the papers, I

62
00:03:30,800 --> 00:03:33,120
see stuff.
And you're the sort of person

63
00:03:33,120 --> 00:03:37,440
where when I read the paper, I
was like, yeah, I really should

64
00:03:37,440 --> 00:03:40,520
have thought of that.
That's a really good idea.

65
00:03:41,680 --> 00:03:45,160
So, yeah, thank you.
No thanks for having me.

66
00:03:45,160 --> 00:03:47,440
It's a real pleasure.
And yeah, looking forward to the

67
00:03:47,480 --> 00:03:51,920
conversation.
So maybe we could get straight

68
00:03:51,920 --> 00:03:56,720
into it, which is the, I guess
the hot conversation now.

69
00:03:57,240 --> 00:04:03,240
Which has only got more intense.
Is this foundation models?

70
00:04:03,280 --> 00:04:05,680
You know, can we somehow make a
model that could?

71
00:04:06,720 --> 00:04:10,400
You know, compute any fluid
flow, whether it's, you know, a

72
00:04:10,400 --> 00:04:12,520
geometry variation, A boundary
condition.

73
00:04:14,040 --> 00:04:20,079
Where do you think we are along
that journey and how close?

74
00:04:20,480 --> 00:04:22,520
How close do you think we are?
To to get in there.

75
00:04:23,880 --> 00:04:28,440
Yeah, I think few years ago I
would have probably given a

76
00:04:28,440 --> 00:04:33,440
different answer, but things are
changing quickly in a in a good

77
00:04:33,440 --> 00:04:36,320
way.
And I think we're getting closer

78
00:04:36,320 --> 00:04:39,480
to having and it all depends on
the type of question that you

79
00:04:39,480 --> 00:04:41,480
want to answer, right.
What type of predictions and

80
00:04:41,480 --> 00:04:45,000
what level of accuracy do you
want to use the systems for

81
00:04:45,000 --> 00:04:48,160
design, Do you want to use the
systems for, for scientific

82
00:04:48,280 --> 00:04:52,040
insight?
And of course what quantities do

83
00:04:52,040 --> 00:04:54,120
you want to get with what level
of accuracy?

84
00:04:54,320 --> 00:04:57,720
So there's many possible
questions and directions to go

85
00:04:57,720 --> 00:05:00,080
into.
But in general, I think that

86
00:05:00,080 --> 00:05:03,640
we're getting pretty close to
having systems that can perform

87
00:05:03,640 --> 00:05:08,200
very well even for quite
complicated quantities and quite

88
00:05:08,200 --> 00:05:11,640
some nuanced phenomena on on our
system.

89
00:05:13,480 --> 00:05:19,280
Especially because depending on
how you define your model, and

90
00:05:19,280 --> 00:05:22,160
depending on the type of
question that you may be

91
00:05:22,160 --> 00:05:27,080
interested in answering, you
don't really need to mimic all

92
00:05:27,080 --> 00:05:31,800
the physics of the renal system.
You might just focus on a subset

93
00:05:31,800 --> 00:05:34,840
of questions or a subset of
mechanisms that could be

94
00:05:35,520 --> 00:05:38,320
represented encapsulated in a
smart way.

95
00:05:38,960 --> 00:05:42,320
And that's basically what your
model does, and not necessarily

96
00:05:42,320 --> 00:05:46,760
all the intricate interscale
mechanisms that give rise to

97
00:05:46,760 --> 00:05:51,640
those particular phenomena.
So yeah, I think we're getting

98
00:05:51,840 --> 00:05:54,840
closer to having systems that
truly can give us the right

99
00:05:54,840 --> 00:05:58,440
accuracy, the right performance
in in quite challenging

100
00:05:58,440 --> 00:05:59,040
problems.
Yeah.

101
00:06:00,320 --> 00:06:05,280
So I'm intrigued to know what's
changed your mind.

102
00:06:05,760 --> 00:06:08,920
We said a few years ago you
would have given a different

103
00:06:08,920 --> 00:06:11,160
handset.
Can you pinpoint anything that

104
00:06:11,160 --> 00:06:15,360
has, yeah, changed over the past
few years that's changed that?

105
00:06:16,000 --> 00:06:20,800
There's a couple of things.
The first one is we realised

106
00:06:20,800 --> 00:06:27,320
that we could identify those key
mechanisms pretty well in quite

107
00:06:27,320 --> 00:06:31,280
complex systems using causality,
explainability.

108
00:06:32,200 --> 00:06:38,040
So not just in idealised
versions of the whole flow or in

109
00:06:38,080 --> 00:06:43,040
reduce order representations,
but in really full fidelity high

110
00:06:43,040 --> 00:06:45,680
order versions of the system, we
could really interrogate the

111
00:06:45,680 --> 00:06:48,560
data, identify what are the
mechanisms that are the most

112
00:06:48,560 --> 00:06:52,400
important and focus on those for
control, for modelling.

113
00:06:52,800 --> 00:06:58,680
So I think those almost
surprising capabilities of this

114
00:06:58,680 --> 00:07:01,440
explain ability and causal
frameworks to really

115
00:07:01,440 --> 00:07:04,920
characterize the systems has
been one big pillar.

116
00:07:05,400 --> 00:07:09,120
And the other one, the
performance of some of the

117
00:07:09,160 --> 00:07:13,040
generative deep learning
methods, especially in the last

118
00:07:13,040 --> 00:07:16,280
few years, diffusion, flow
matching, conditional later

119
00:07:16,280 --> 00:07:21,080
diffusion, these systems are
really capable of generalizing

120
00:07:21,120 --> 00:07:25,800
to an extent that I mean not so
long ago I would say that it was

121
00:07:25,800 --> 00:07:29,120
not really possible.
There's still quite some more to

122
00:07:29,120 --> 00:07:32,360
do and there needs to be some
smart way of formulating these

123
00:07:32,360 --> 00:07:37,040
problems in order to achieve
real generalization, but I would

124
00:07:37,040 --> 00:07:41,400
say that that we are really in a
stage where things are pretty

125
00:07:41,400 --> 00:07:44,400
impressive.
Maybe you could explain, pardon

126
00:07:44,400 --> 00:07:47,720
the pun, what you mean by
explainable AI.

127
00:07:48,680 --> 00:07:51,120
Yeah, that's a, that's a good a
good question.

128
00:07:51,520 --> 00:07:56,840
So what we mean with the systems
and what constitutes an

129
00:07:56,840 --> 00:08:02,440
explanation and even the, the,
you know, the, the grammar and

130
00:08:02,440 --> 00:08:05,240
the definition changes a little
bit depending on the, on the

131
00:08:05,360 --> 00:08:09,600
field you have explainable AI,
explainable deep learning

132
00:08:09,600 --> 00:08:12,640
interpretability.
So there's different kind of

133
00:08:12,640 --> 00:08:17,880
flavours to it.
But the way that I use this term

134
00:08:17,880 --> 00:08:23,000
explainable deep learning would
be kind of assessing for a

135
00:08:23,000 --> 00:08:28,360
particular model what features
of your input matter the most to

136
00:08:28,480 --> 00:08:33,159
the prediction of that model.
So really identifying kind of

137
00:08:33,159 --> 00:08:38,840
like in a feature attribution
sense, what, what, which ones of

138
00:08:38,840 --> 00:08:41,880
those components of the input
really matter to your output.

139
00:08:42,480 --> 00:08:45,840
And why is that interest in the
context of high dimensional

140
00:08:46,040 --> 00:08:49,880
chaotic engineering systems?
Because if you create a model,

141
00:08:49,880 --> 00:08:53,360
an input output model, where I'm
just going to give you an

142
00:08:53,360 --> 00:08:57,360
example.
If I have the wing of an

143
00:08:57,360 --> 00:09:01,800
aircraft and I really want to
see what flow motions are really

144
00:09:01,800 --> 00:09:06,560
affecting the most the drag,
well, then I can create a model

145
00:09:06,560 --> 00:09:10,080
which given the flow, it just
predicts the drag on the wing.

146
00:09:10,840 --> 00:09:13,840
And that's a predictive model
that hopefully performs

147
00:09:13,840 --> 00:09:17,320
reasonably well.
But if I use this explainability

148
00:09:17,320 --> 00:09:22,720
methods on that model, then I
can go point by point and really

149
00:09:22,720 --> 00:09:25,720
identify which points are
contributing the most to the

150
00:09:25,720 --> 00:09:30,480
drag on that wing.
And that idea can really allow

151
00:09:30,480 --> 00:09:35,440
us to identify volumes of
importance of importance for

152
00:09:35,440 --> 00:09:37,960
that particular question, for
the drag in this case.

153
00:09:38,360 --> 00:09:41,680
And what we have found is 2
interesting things.

154
00:09:42,080 --> 00:09:46,560
One is by using these
explainability tools, we realize

155
00:09:46,640 --> 00:09:52,640
that the classical approaches to
study turbulence and people have

156
00:09:52,640 --> 00:09:57,160
been looking at vortices,
streaks, Rhino stresses, many

157
00:09:57,160 --> 00:10:01,360
other quantities, right?
In fact, when people want to

158
00:10:01,360 --> 00:10:06,160
show off a bit and they do a
very nice CFD visualisation,

159
00:10:06,840 --> 00:10:09,840
very colourful visualisation,
they will show you Lambda 2,

160
00:10:09,840 --> 00:10:12,280
right, or something similar.
They show you vortices because

161
00:10:12,280 --> 00:10:14,000
it's very spectacular and it's
very turbulent.

162
00:10:14,480 --> 00:10:18,280
And what we realise with our
methods is that the vortices

163
00:10:18,480 --> 00:10:21,920
don't matter so much actually,
at least for the friction and

164
00:10:21,920 --> 00:10:23,800
for the drug.
They matter for other things,

165
00:10:23,800 --> 00:10:27,320
but not for the stuff that you
are trying to show them for.

166
00:10:27,760 --> 00:10:31,760
And the reason is the fact that
you can see something does not

167
00:10:31,760 --> 00:10:34,600
mean that it's important.
You simply can't see it.

168
00:10:34,760 --> 00:10:39,480
And a lot of the turbulence, a
lot of the turbulence research

169
00:10:39,480 --> 00:10:43,400
has been a bit misled by what
you could see in experiments,

170
00:10:43,400 --> 00:10:46,080
which of course it has been very
helpful and it has been very

171
00:10:46,360 --> 00:10:50,200
illustrative.
But sometimes you need to step

172
00:10:50,200 --> 00:10:54,040
away a bit from the physical
intuition and just simply

173
00:10:54,040 --> 00:10:56,680
interrogate the data in a more
diagnostic way and let the data

174
00:10:56,680 --> 00:10:59,000
tell you what matters and what
doesn't matter.

175
00:10:59,560 --> 00:11:03,480
So what we've found is the
classical approaches to

176
00:11:04,280 --> 00:11:07,440
turbulence research.
They were only telling a part of

177
00:11:07,440 --> 00:11:11,560
the story.
So in different regions, close

178
00:11:11,560 --> 00:11:15,000
to the wall, the the Renaissance
stresses were important.

179
00:11:15,160 --> 00:11:16,960
Farther away, the streets were
important.

180
00:11:17,080 --> 00:11:19,240
Farther away there were all the
rhino stresses that were

181
00:11:19,240 --> 00:11:22,840
important.
So the classical views were not

182
00:11:23,040 --> 00:11:26,080
wrong because obviously, you
know, there's a lot of time and

183
00:11:26,080 --> 00:11:27,080
effort.
They brought it to it.

184
00:11:27,360 --> 00:11:29,160
They were just telling part of
the story, right?

185
00:11:30,160 --> 00:11:31,880
Like the three blind men and the
elephant.

186
00:11:32,040 --> 00:11:36,040
They were, you know, giving
different partial views of what

187
00:11:36,040 --> 00:11:39,080
the elephant look like.
But all the classical

188
00:11:39,080 --> 00:11:42,680
perspectives together is what
actually comes close to what we

189
00:11:42,680 --> 00:11:45,440
identify with our explainability
methods.

190
00:11:46,280 --> 00:11:49,760
And why are we so sure that
these methods are giving us

191
00:11:49,760 --> 00:11:51,800
something that is interesting
and physical?

192
00:11:52,080 --> 00:11:55,560
Because when we device control
mechanisms, when we manipulate

193
00:11:55,560 --> 00:12:00,120
the flow to diminish the
presence of these mechanisms

194
00:12:00,120 --> 00:12:03,000
that we identify, that's when we
get the highest track reduction.

195
00:12:03,720 --> 00:12:09,880
So it's not just that purely we
can find nice colourful volumes

196
00:12:09,880 --> 00:12:13,840
of extractors is that those
extractors in a causal way are

197
00:12:13,880 --> 00:12:16,960
actually affecting the drug in
this case the most.

198
00:12:17,280 --> 00:12:23,120
And they are tackling those
extractors is really looking at

199
00:12:23,120 --> 00:12:28,480
the root cause of the drug, so
the disease and not the symptom.

200
00:12:29,000 --> 00:12:32,760
And that's why we think that
this is really a good way to

201
00:12:32,760 --> 00:12:38,120
look at these mechanisms and not
only from kind of controlled

202
00:12:38,120 --> 00:12:42,320
perspective or from a
satisfactory knowledge point of

203
00:12:42,320 --> 00:12:45,600
view, also from a modelling
point of view, because those

204
00:12:45,600 --> 00:12:49,680
mechanisms are the ones that
contain the crucial information

205
00:12:49,680 --> 00:12:52,240
if you want to be a good model
for that phenomenon.

206
00:12:53,800 --> 00:13:02,120
So how do you see the the
challenge that I still be not

207
00:13:02,120 --> 00:13:04,280
got my head around which is the
following.

208
00:13:05,160 --> 00:13:09,280
If we assume that for a
foundational model or you know,

209
00:13:09,360 --> 00:13:14,120
surrogate model, we're assuming
that we have lots of data of

210
00:13:14,200 --> 00:13:17,040
wings of cars, of buildings,
etcetera.

211
00:13:18,160 --> 00:13:23,440
At least today, you know, most
of those have, you know, a

212
00:13:23,440 --> 00:13:27,680
decent error to what is the real
turbulence, what is the real

213
00:13:27,680 --> 00:13:30,880
flow?
And you're asking the model to

214
00:13:30,880 --> 00:13:33,040
learn it.
And sometimes the data comes

215
00:13:33,040 --> 00:13:36,960
from different sources.
You know, the wing may have been

216
00:13:36,960 --> 00:13:39,560
done with this turbulence model,
The car may have done with this.

217
00:13:39,560 --> 00:13:45,960
So how do you see the foundation
model?

218
00:13:46,960 --> 00:13:53,080
Is it ultimately yeah.
Do we have to train on DNS?

219
00:13:53,080 --> 00:13:55,480
But then we can't train on DNS
because it will just be too

220
00:13:55,480 --> 00:13:57,360
expensive.
Do you need some DNSI?

221
00:13:57,680 --> 00:14:03,920
Still feel like the foundation
model is ultimately going to not

222
00:14:03,920 --> 00:14:07,680
be as good as people think.
Yeah, because no, I totally see

223
00:14:07,680 --> 00:14:11,920
your your point.
And I think here the key is

224
00:14:11,920 --> 00:14:15,720
going back a bit to the one of
the first questions, what do we

225
00:14:15,720 --> 00:14:21,040
want this model for, right?
And I mean, I don't think that

226
00:14:21,040 --> 00:14:24,760
we are trying to build these
models to replace DNS, right?

227
00:14:24,760 --> 00:14:28,280
I mean we're not going to
because that computationally

228
00:14:28,280 --> 00:14:30,240
would not really make sense,
right?

229
00:14:30,560 --> 00:14:34,240
So I think that we're trying to
build these models to either

230
00:14:34,240 --> 00:14:39,840
accelerate design and
optimization and or achieve some

231
00:14:39,840 --> 00:14:42,520
sort of scientific insight in
our system.

232
00:14:43,000 --> 00:14:49,320
So how can we handle data from
different fidelities and

233
00:14:49,320 --> 00:14:53,120
different modalities?
That's something that in what

234
00:14:53,120 --> 00:14:56,120
we're building in my group,
we're trying to embed the

235
00:14:56,280 --> 00:14:59,800
uncertainty quantification with
active learning loops in a sense

236
00:14:59,800 --> 00:15:03,080
that we know that not all all
our data is of the same

237
00:15:03,080 --> 00:15:08,000
fidelity.
When we are trying to generalize

238
00:15:08,000 --> 00:15:12,560
and we're trying to produce data
for new cases, the system will

239
00:15:12,560 --> 00:15:14,720
tell us, look for this
particular case that you're

240
00:15:14,720 --> 00:15:19,760
trying to produce data in some
latent representation, your

241
00:15:19,760 --> 00:15:23,000
uncertainty is very high.
So you should probably go back

242
00:15:23,000 --> 00:15:26,640
and if you know, if you're
trying to do a helicopters,

243
00:15:26,640 --> 00:15:29,000
well, your helicopter data is
actually pretty bad.

244
00:15:29,000 --> 00:15:32,680
So you should be there, have
higher fidelity or, you know,

245
00:15:32,680 --> 00:15:35,440
run more experiments or really
improve your model there.

246
00:15:35,680 --> 00:15:38,520
And I think that's one strength
actually that we can

247
00:15:38,680 --> 00:15:42,240
progressively keep improving our
model as new data becomes a

248
00:15:42,240 --> 00:15:47,280
while.
Now your data for helicopters is

249
00:15:47,280 --> 00:15:50,080
pretty bad for what, right,
Because it depends on what

250
00:15:50,080 --> 00:15:53,360
you're trying to do.
And here is when we need to be a

251
00:15:53,360 --> 00:15:58,360
bit a bit realistic about the
metrics and about the targets

252
00:15:58,360 --> 00:16:02,440
that we want to hit.
So if we are trying to get lift

253
00:16:02,440 --> 00:16:06,760
and drag, right, that's one type
of accuracy and one type of

254
00:16:06,760 --> 00:16:09,200
model.
If we're trying to get the

255
00:16:09,200 --> 00:16:12,800
spectrum right and all the
interscale mechanisms, right,

256
00:16:12,800 --> 00:16:14,360
then that's a different type of
model, right?

257
00:16:14,720 --> 00:16:17,640
So I think in that sense,
depending on the type of

258
00:16:17,920 --> 00:16:19,840
approach and the type of
figurative that we want to

259
00:16:19,840 --> 00:16:24,000
achieve, we may have to resort
to different data sets and

260
00:16:24,000 --> 00:16:26,560
different ways of assessing the
error and the uncertainty of our

261
00:16:26,560 --> 00:16:30,560
systems.
Yeah, no, I, I agree.

262
00:16:30,560 --> 00:16:34,760
I think the there'll be probably
one side of the community that

263
00:16:35,520 --> 00:16:40,920
happily uses, you know, Reynolds
averaged or even panel methods

264
00:16:40,920 --> 00:16:44,440
or or or lower and are, you
know, quite comfortable with

265
00:16:44,440 --> 00:16:47,960
models that are not of DNS
accuracy.

266
00:16:48,400 --> 00:16:53,280
But yeah, yeah.
On the other hand, I, I, I can

267
00:16:53,280 --> 00:16:57,160
imagine just as companies spend
a fortune constantly trying to

268
00:16:57,160 --> 00:17:00,160
get to higher and higher
fidelity CFD that maybe the

269
00:17:00,160 --> 00:17:02,440
models will, you know, go with
it.

270
00:17:03,480 --> 00:17:06,480
One question, because I think
some of your background, you

271
00:17:06,480 --> 00:17:11,760
know, was on, I guess, would it
be fair to say like reduced

272
00:17:11,760 --> 00:17:18,440
order modelling and and some of
that side 2, maybe a more like

273
00:17:19,240 --> 00:17:20,680
lay audience?
I'm not saying everyone

274
00:17:20,680 --> 00:17:21,920
listening to this is a lay
audience.

275
00:17:21,920 --> 00:17:23,839
That's quite the opposite.
But there are some people who

276
00:17:23,839 --> 00:17:29,640
are maybe not as grounded in all
the latest common question I get

277
00:17:29,640 --> 00:17:32,880
is, well, machine learning is
just reduced order modelling.

278
00:17:34,320 --> 00:17:38,760
How would you chart the
evolution, you know, from like

279
00:17:40,160 --> 00:17:43,680
what people would call ROMs
before to machine learning?

280
00:17:43,680 --> 00:17:47,800
And what is it about modern
machine learning that separates

281
00:17:47,800 --> 00:17:52,920
it from reduced automobile?
Unless you classify them as to

282
00:17:52,920 --> 00:17:53,600
say.
Yeah.

283
00:17:53,840 --> 00:17:58,640
No, that's a good question.
Depends a little bit on what

284
00:17:58,640 --> 00:18:01,520
you're trying to do with your
machine learning, right, Because

285
00:18:01,520 --> 00:18:04,400
also machine learning is a quite
broad area.

286
00:18:05,080 --> 00:18:11,120
So, I mean, and we have some
work that we've done before on

287
00:18:11,120 --> 00:18:14,880
machine learning for CFD.
So how should you or what are

288
00:18:14,880 --> 00:18:17,680
the areas where machine learning
can help CFD?

289
00:18:18,080 --> 00:18:21,000
And this is work that I did with
my friend Steve Branton some

290
00:18:21,000 --> 00:18:23,880
years ago where we found three
areas.

291
00:18:23,880 --> 00:18:28,520
We found accelerate DNS,
improved models, both runs and

292
00:18:28,520 --> 00:18:32,960
LES and improve reduce order
models or saturate models.

293
00:18:33,120 --> 00:18:36,360
And, and, and, and that's if we
are thinking of machine learning

294
00:18:36,360 --> 00:18:40,560
for CFD, something that we made
quite clear at the time, this is

295
00:18:40,560 --> 00:18:45,200
4 years ago was that machine
learning in principle would not

296
00:18:45,200 --> 00:18:48,200
replace CFD.
It's not about, and I think that

297
00:18:48,200 --> 00:18:50,720
many people when they think
machine learning for CFD,

298
00:18:50,880 --> 00:18:53,280
they're thinking automatically
replacing CFD, right.

299
00:18:53,640 --> 00:18:56,440
I, I don't think that it's
really about replacing CFD, but

300
00:18:56,440 --> 00:18:58,320
rather complementing and
helping.

301
00:18:58,680 --> 00:19:03,760
So reduce order modelling would
be just one area within all the

302
00:19:03,760 --> 00:19:07,880
spectrum of possibilities for
machine learning can help just

303
00:19:07,880 --> 00:19:10,720
within CFD, but we can also
think about control and

304
00:19:10,720 --> 00:19:14,720
optimization where this
fantastic reinforcement learning

305
00:19:14,720 --> 00:19:17,640
work, for example, on really
finding new controller

306
00:19:17,640 --> 00:19:21,880
strategies.
So it's very broad, but I would

307
00:19:21,880 --> 00:19:26,480
say that if we focus on on
surrogate models, traditional

308
00:19:26,480 --> 00:19:34,120
models, one key has been the
possibility of hiring well,

309
00:19:34,120 --> 00:19:36,800
nonlinearities in your in your
model development, right in your

310
00:19:36,800 --> 00:19:41,960
surrogate.
And of course PODDMDI mean the

311
00:19:42,320 --> 00:19:45,080
mostly linear, although you can
have nonlinearities in there.

312
00:19:45,560 --> 00:19:50,960
But having neural network based
compression at scale, which is

313
00:19:50,960 --> 00:19:53,600
what you can have now with more,
more than out in core

314
00:19:53,600 --> 00:19:58,440
architectures that really allows
you to have very impressive

315
00:19:58,440 --> 00:20:02,400
compression rates, right.
So quite some capability to to

316
00:20:02,640 --> 00:20:05,520
distill the essential physics of
your system.

317
00:20:06,160 --> 00:20:10,640
And something that we found also
in some of our work is in, I

318
00:20:10,640 --> 00:20:12,680
mean, out on corners are good at
compressing, right?

319
00:20:12,760 --> 00:20:16,400
Compressing anything, images of
cats and dogs and images of tool

320
00:20:16,400 --> 00:20:20,160
and flows if you wish.
But tool and flows, their

321
00:20:20,160 --> 00:20:23,480
images, their flow fields, they
contain much more information.

322
00:20:23,480 --> 00:20:25,960
They're much richer than cats
and dogs, right?

323
00:20:26,080 --> 00:20:27,880
There's a spectrum, there's a
range of scales.

324
00:20:28,240 --> 00:20:32,840
So you should somehow, when you
compress, you should somehow

325
00:20:32,840 --> 00:20:36,360
disentangle the latent
representation, because if you

326
00:20:36,360 --> 00:20:39,640
do that and you can use beta
AES, you can use hierarchical

327
00:20:39,640 --> 00:20:41,720
priors.
There's different ways to do

328
00:20:42,080 --> 00:20:46,840
that disentanglement, but that
really allows you to encapsulate

329
00:20:46,840 --> 00:20:49,160
in different latent
representations and different

330
00:20:49,160 --> 00:20:51,800
latent vectors, different
physical phenomena of your

331
00:20:51,800 --> 00:20:54,960
problem.
And that's going to be more

332
00:20:54,960 --> 00:20:57,600
effective from the
interpretability point of view

333
00:20:58,120 --> 00:21:01,640
of your Stargate system, but
also from a predictive point of

334
00:21:01,640 --> 00:21:03,320
view.
If you want to make predictions

335
00:21:03,320 --> 00:21:06,680
in in the latent space, that
disentanglement is actually

336
00:21:06,680 --> 00:21:08,560
going to be quite, quite
helpful.

337
00:21:08,960 --> 00:21:13,200
So, yeah, I mean, I guess it was
a bit of a long answer to your

338
00:21:13,200 --> 00:21:15,800
question, but machine learning
is not just reduce order

339
00:21:15,800 --> 00:21:17,200
modelling.
Let's say that there's many

340
00:21:17,240 --> 00:21:21,080
other things that one can do and
within reduce order modelling

341
00:21:21,080 --> 00:21:25,920
that, you know, capability of
really having out in colors at

342
00:21:25,920 --> 00:21:29,240
scale and disentangling the
latent representations has been

343
00:21:29,240 --> 00:21:32,480
quite critical.
And also Transformers, that's

344
00:21:32,480 --> 00:21:37,400
another, well dimension let's
say, or another topic, which

345
00:21:37,400 --> 00:21:40,280
also has helped in building the
temporal dynamics of these radio

346
00:21:40,280 --> 00:21:43,440
solar systems.
I mean, where do you see?

347
00:21:43,840 --> 00:21:50,760
At least my impression has been
that even though the turbulent

348
00:21:50,760 --> 00:21:57,040
modelling started off as being
one of the the big focuses of

349
00:21:57,040 --> 00:22:02,480
the fluent community, it does
seem to have slightly petered

350
00:22:02,480 --> 00:22:07,400
out a little bit and the focus
has shifted a little bit more to

351
00:22:07,400 --> 00:22:09,920
the surrogate modelling, to the
sort of reduced auto modelling.

352
00:22:10,280 --> 00:22:12,640
Is that something that you've
also noticed?

353
00:22:12,640 --> 00:22:16,520
And do you think that's just
because of a lack of progress?

354
00:22:16,720 --> 00:22:21,280
Do you think it's because the,
you know, the improvements?

355
00:22:21,280 --> 00:22:24,760
I'm just wondering if how you've
seen that?

356
00:22:25,560 --> 00:22:29,000
Yeah, no, that's another good
point.

357
00:22:30,600 --> 00:22:36,040
I think that so turbulence
modelling can be done in

358
00:22:36,040 --> 00:22:40,680
different ways.
And One Direction that has been

359
00:22:40,680 --> 00:22:46,760
adopted by many people has been
to start where more classical

360
00:22:46,760 --> 00:22:49,880
approaches to turbulence model
modelling start.

361
00:22:49,880 --> 00:22:54,680
Basically to which in a way is
just making an empirical

362
00:22:54,680 --> 00:22:57,680
assumption of how the, you know,
the Renaissance stresses need to

363
00:22:57,680 --> 00:22:59,680
behave.
To some extent.

364
00:22:59,680 --> 00:23:01,720
There is some empirical
assumption, more or less

365
00:23:01,720 --> 00:23:06,240
sophisticated and then try to
use machine learning systems to

366
00:23:06,240 --> 00:23:10,200
continue from there.
So in, in other words, to fit

367
00:23:10,200 --> 00:23:12,560
coefficients on classical
models, right?

368
00:23:13,040 --> 00:23:18,000
And well, that's not maybe the
most revolutionary thing, right?

369
00:23:18,040 --> 00:23:22,600
Because if you do that, and
that's what many people have

370
00:23:22,640 --> 00:23:24,480
done.
And I think that's why a bit of

371
00:23:24,480 --> 00:23:28,840
the ML community could have got
a bit disappointed with ML for

372
00:23:28,840 --> 00:23:33,880
fluids, because they were using
ML to adjust coefficients in two

373
00:23:33,880 --> 00:23:36,600
large modelling models that have
been around for decades and

374
00:23:36,600 --> 00:23:40,440
which have based on assumptions
that are a bit crude, perhaps,

375
00:23:40,440 --> 00:23:42,240
right.
But for a reason, they were

376
00:23:42,240 --> 00:23:46,320
crude because there were not
other approaches to, to

377
00:23:46,320 --> 00:23:49,760
modelling that were, you know,
kept possible at that time.

378
00:23:51,240 --> 00:23:54,720
So I think that could be why
there has been a bit of slow

379
00:23:54,720 --> 00:23:57,280
down.
But at the same time, I've seen

380
00:23:57,280 --> 00:24:01,800
some promising approaches.
I mean, I'm, I'm a bit of a fan

381
00:24:01,800 --> 00:24:04,440
of reinforcement learning.
I've, I've been developing many

382
00:24:04,440 --> 00:24:07,760
methods within reinforcement
learning for control, for

383
00:24:07,760 --> 00:24:10,800
optimization and also for
modelling.

384
00:24:10,800 --> 00:24:13,800
We have some, some projects on
reinforcement learning for

385
00:24:13,800 --> 00:24:16,880
modelling and there's other
groups who are doing great

386
00:24:16,880 --> 00:24:24,040
things in this space.
And I think that the idea is why

387
00:24:25,400 --> 00:24:30,600
start with a model that has been
around for, you know, 20305060

388
00:24:30,600 --> 00:24:35,960
years and try to adjust those
coefficients when we can maybe

389
00:24:35,960 --> 00:24:38,360
take one step back.
So instead of asking my machine

390
00:24:38,360 --> 00:24:42,760
learning model to say for the
Smolensky model, what should be

391
00:24:42,760 --> 00:24:48,400
my coefficient, Can you find it
to, to, or even say to my

392
00:24:48,400 --> 00:24:54,080
machine learning model or just
find the subway scale tensor to

393
00:24:54,080 --> 00:24:58,520
be able to fit the statistics of
this this particular channel or

394
00:24:58,760 --> 00:25:01,800
particular flow?
Maybe we can be a bit more

395
00:25:02,720 --> 00:25:07,320
abstract and say, look, I don't
know what the mean flow or the

396
00:25:07,320 --> 00:25:10,000
fluctuations will look like in
this particular case.

397
00:25:10,080 --> 00:25:13,280
I, I don't know what happened.
I've never run this wing or this

398
00:25:13,280 --> 00:25:18,000
aircraft, but what I know is
that turbulence is characterized

399
00:25:18,000 --> 00:25:21,240
by certain energy transfer
mechanisms across scales, and

400
00:25:21,240 --> 00:25:24,000
that energy goes in this
election and there is an inverse

401
00:25:24,000 --> 00:25:26,280
cascade.
And there's certain phenomena

402
00:25:26,680 --> 00:25:29,440
that need to be true from a
spectral point of view.

403
00:25:29,440 --> 00:25:34,160
And from a mechanistic point of
view, whatever your model is and

404
00:25:34,160 --> 00:25:38,600
whatever the forcing term that
you get each step from your

405
00:25:38,600 --> 00:25:40,720
reinforcement learning or
whatever optimizer that you're

406
00:25:40,720 --> 00:25:43,680
using, just respect those
physical constraints.

407
00:25:44,000 --> 00:25:46,560
Because my flow needs to be
physically correct.

408
00:25:47,120 --> 00:25:49,880
And physically correct does not
mean this is your mean flow.

409
00:25:49,880 --> 00:25:52,600
You need to adjust to this mean
flow because that's a circular

410
00:25:52,600 --> 00:25:53,960
argument, then you need the mean
flow, right?

411
00:25:54,480 --> 00:25:59,720
But rather not only a circular
argument, it's some more, it's a

412
00:25:59,720 --> 00:26:04,040
much more constraining goal to
fit the statistics right.

413
00:26:04,360 --> 00:26:09,680
But something a bit more
flexible, such as making sure

414
00:26:09,680 --> 00:26:13,480
that the energy fluxes are
correct step by step is

415
00:26:13,480 --> 00:26:15,720
something that is less impose
imposing.

416
00:26:16,400 --> 00:26:21,480
And it's also something that is
a bit more manageable for a for

417
00:26:21,480 --> 00:26:25,040
a system for an optimization
system to to be able to adapt

418
00:26:25,040 --> 00:26:27,440
step by step such that those
fluxes are OK.

419
00:26:27,760 --> 00:26:31,680
So I think the second direction
of making statements that are a

420
00:26:31,680 --> 00:26:36,520
bit more general and still
correct from a physical point of

421
00:26:36,520 --> 00:26:39,600
view, that could be more
promising to develop more.

422
00:26:41,880 --> 00:26:46,400
I mean, one of the things that
I've started contemplating, and

423
00:26:46,400 --> 00:26:51,080
I'm interested to see if you've
been the same is for for a while

424
00:26:51,080 --> 00:26:56,640
I was only thinking about, you
know, the machine learning,

425
00:26:57,280 --> 00:27:01,560
either in the context of
learning what the total and

426
00:27:01,560 --> 00:27:04,960
viscosity should be or, or more
broadly, what a surrogate model,

427
00:27:05,480 --> 00:27:09,440
you know, should be in terms of
a transformer based or some

428
00:27:09,440 --> 00:27:13,480
neural operator or, or whatever.
And the criticism has been that

429
00:27:13,480 --> 00:27:20,440
it's, how can it match the, the
brains, you know, the thought

430
00:27:20,440 --> 00:27:22,400
process that, that, that we've
had.

431
00:27:22,400 --> 00:27:26,960
And I saw at an event, I was at
the British Library, there was

432
00:27:27,680 --> 00:27:32,720
an AI, somebody organized UKTC,
the turbans consortium in the

433
00:27:32,720 --> 00:27:36,440
UK.
And it was Luca Magri from

434
00:27:36,440 --> 00:27:38,600
Imperial was organizing with
some people.

435
00:27:39,160 --> 00:27:42,720
And there was a Professor,
Michael Leschiner, who, you

436
00:27:42,720 --> 00:27:46,160
know, is a sort of pioneer of
turbans modelling developed with

437
00:27:46,160 --> 00:27:48,280
Brian Launder, you know, all
these renal stress models.

438
00:27:48,680 --> 00:27:52,840
And, and he rightly said to me
that he was, you know, gently

439
00:27:52,840 --> 00:27:58,560
wondering how could these AI
models, you know, know, all

440
00:27:58,560 --> 00:28:00,720
these things that, that, that
they've done.

441
00:28:01,440 --> 00:28:05,640
And then I started to think, is
this really, and I don't know if

442
00:28:05,640 --> 00:28:09,720
the capabilities today where the
agentic thing comes in because

443
00:28:09,720 --> 00:28:16,920
in some ways is it more
realistic to ask an LLM to

444
00:28:16,920 --> 00:28:21,960
reason through the process of
how you would develop a turbans

445
00:28:21,960 --> 00:28:29,400
model and have links in to
software tools to explore it and

446
00:28:29,400 --> 00:28:31,880
go through that reasoning
process and reinforcement

447
00:28:31,880 --> 00:28:34,520
learning.
Then just say, hey, try and find

448
00:28:35,240 --> 00:28:38,120
me the the latent representation
of this.

449
00:28:38,200 --> 00:28:39,880
Do you know what I mean?
Do you think almost the

450
00:28:39,880 --> 00:28:45,760
scientific discovery that piece
is the better route rather than

451
00:28:45,760 --> 00:28:48,840
just expecting machine that is?
So I don't know if I've

452
00:28:48,840 --> 00:28:50,960
explained that, but.
That makes perfect sense.

453
00:28:50,960 --> 00:28:55,640
I think it's, it's a very fair
question and we are trying to

454
00:28:56,240 --> 00:28:59,320
well to really first of all
understand the potential of

455
00:28:59,320 --> 00:29:05,160
decisioning systems and 2nd, we
are learning to ask the right

456
00:29:05,160 --> 00:29:10,040
questions to these systems.
I think before getting into

457
00:29:10,040 --> 00:29:15,320
that, I think that a preliminary
point is what is the right way

458
00:29:15,320 --> 00:29:21,800
to represent the data that we as
engineers and scientists in you

459
00:29:21,800 --> 00:29:23,960
know, higher order chaotic
complex systems.

460
00:29:24,160 --> 00:29:27,040
What is the right, the right way
to represent the data and such

461
00:29:27,080 --> 00:29:31,200
an exploration from these agents
or whatever or, or any

462
00:29:31,200 --> 00:29:35,160
scientist, what's the right
platform to represent the data,

463
00:29:35,160 --> 00:29:37,960
right?
Because there's quite some work

464
00:29:37,960 --> 00:29:42,120
on LLMS to do this and to try to
achieve discovery, whatever we

465
00:29:42,120 --> 00:29:47,160
define by discovery.
But you know, if we think of the

466
00:29:48,000 --> 00:29:52,360
of turbulent flows, of course,
this is these are systems that

467
00:29:52,360 --> 00:29:55,440
are having the broadband in
terms of the spectrum, they have

468
00:29:55,440 --> 00:29:57,800
multiple skills, multiple energy
fluxes.

469
00:29:58,240 --> 00:30:00,520
Perhaps text is not the right
way to represent this very

470
00:30:00,520 --> 00:30:06,040
complex data, right?
And perhaps using LMS directly

471
00:30:06,400 --> 00:30:10,360
and text as a platform to try to
explore complex questions in

472
00:30:10,360 --> 00:30:15,200
this space might not be the the
most suitable approach.

473
00:30:15,640 --> 00:30:19,880
And what we have been thinking
and exploring is what if we can

474
00:30:19,880 --> 00:30:25,480
create latent representations
that are particularly attuned to

475
00:30:26,160 --> 00:30:28,920
represent important properties
of these two water flow systems.

476
00:30:29,680 --> 00:30:32,240
And, and that's an approach that
we have been developing in our

477
00:30:32,240 --> 00:30:37,000
group where we are trying to
build a foundation models, kind

478
00:30:37,000 --> 00:30:44,320
of latent representations of
very broad ranges of cases such

479
00:30:44,320 --> 00:30:48,400
that we can use those latent
spaces to explore the, the, you

480
00:30:49,160 --> 00:30:51,840
know, the, the complexity of
these, of these two water flows.

481
00:30:52,920 --> 00:30:58,080
One guiding principle has been
information, information fluxes,

482
00:30:58,560 --> 00:31:03,520
causality.
So what if we can create latent

483
00:31:03,520 --> 00:31:07,680
spaces where causal relations
are maximized or where the

484
00:31:07,680 --> 00:31:11,320
disentanglement of the variables
is such that I can very easily

485
00:31:11,320 --> 00:31:13,960
explore physical mechanisms
through that latent space.

486
00:31:14,440 --> 00:31:18,600
And I believe and again, this is
stuff that we are very excited

487
00:31:18,600 --> 00:31:22,400
about because we are we're
developing it with pretty

488
00:31:22,400 --> 00:31:26,880
promising results that using a
genetic systems in that latent

489
00:31:26,880 --> 00:31:31,800
representation might be a good
way to achieve discovery.

490
00:31:32,160 --> 00:31:40,720
So in, in other words, I have a
data from wings aircraft flows

491
00:31:40,720 --> 00:31:44,000
in cities a compressors and
these are very different fluid

492
00:31:44,000 --> 00:31:48,080
mechanics problems.
But if we find smart way to

493
00:31:48,080 --> 00:31:51,920
compress all that data and
express it in the same latent

494
00:31:51,920 --> 00:31:55,680
representation, so now suddenly
I don't have a wing and a city,

495
00:31:55,880 --> 00:31:59,120
but everything is expressed in
the same language on the same

496
00:31:59,120 --> 00:32:02,120
table.
And then I can allow this

497
00:32:02,560 --> 00:32:06,480
agentic systems to explore this
latent space very efficiently.

498
00:32:06,520 --> 00:32:10,400
Because of course this AI
systems can't really find one

499
00:32:10,400 --> 00:32:12,920
side compress everything so
much.

500
00:32:13,360 --> 00:32:17,800
I can find patterns across cases
that might not be obvious in in

501
00:32:17,800 --> 00:32:22,800
the physical space for us humans
and neither for agents, right.

502
00:32:22,800 --> 00:32:26,000
Because finding a connection
between a wing and a city, I

503
00:32:26,000 --> 00:32:29,040
mean, yeah, maybe if I look like
that, I see a vortex that looks

504
00:32:29,040 --> 00:32:30,600
similar.
But you know, it's, it's a bit

505
00:32:30,600 --> 00:32:32,360
more anecdotal, anecdotal than
anything.

506
00:32:32,600 --> 00:32:36,920
But when I express everything in
that space and then I can do

507
00:32:36,920 --> 00:32:40,600
that automatic and autonomous
exploration with these agents,

508
00:32:41,200 --> 00:32:44,640
then I can try to get inside.
Then I can try to find causal

509
00:32:44,640 --> 00:32:46,880
relations that are key in that
later space.

510
00:32:47,480 --> 00:32:52,480
And then our job with
explainability tools is to

511
00:32:52,720 --> 00:32:55,640
express that insight in the
latent space, which is

512
00:32:55,640 --> 00:32:59,840
completely non understandable
for us back in the physical

513
00:32:59,840 --> 00:33:02,880
space and say, look, this is
what insight looks like in the

514
00:33:02,880 --> 00:33:05,800
latent space.
That means that this vortex from

515
00:33:05,800 --> 00:33:08,120
the sheer layer of the wing is
actually very similar to this

516
00:33:08,120 --> 00:33:10,800
vortex from the sheer layer in
the separation of that building

517
00:33:10,800 --> 00:33:13,160
in the city.
So that's actually a common

518
00:33:13,160 --> 00:33:14,920
mechanism.
That's interesting.

519
00:33:15,280 --> 00:33:19,800
And I can find that through
compressing, expressing in the

520
00:33:19,800 --> 00:33:24,080
common space, A autonomously
exploring and then going back to

521
00:33:24,080 --> 00:33:27,800
the physical representation.
Yeah, that's, I really like the

522
00:33:28,600 --> 00:33:31,600
I've you've hit the really good
explanation there because this

523
00:33:31,600 --> 00:33:37,080
is something I've been thinking
in my head is and I I'm not, I

524
00:33:37,080 --> 00:33:39,080
need to look into it more to be
completely honest.

525
00:33:39,080 --> 00:33:42,720
But I almost see 2 slightly
different approaches

526
00:33:42,720 --> 00:33:48,720
traditionally, which is 1, which
is kind of what we wrote in this

527
00:33:48,720 --> 00:33:50,840
paper.
We put out recently this idea

528
00:33:50,840 --> 00:33:53,880
that, oh, you, you know, we're
going to run lots of simulations

529
00:33:53,880 --> 00:33:57,680
of planes and cars and cities
and turbo machineries and

530
00:33:57,680 --> 00:34:01,560
basically every single possible
geometry and band of condition

531
00:34:01,560 --> 00:34:07,240
that we see in real life, which
feels in some ways the more

532
00:34:07,920 --> 00:34:12,080
logical and achievable way.
It's just like you just have to

533
00:34:12,080 --> 00:34:15,159
have a lot of them.
But then the other side is that

534
00:34:16,040 --> 00:34:21,840
actually it doesn't matter
whether it's a plane or a city

535
00:34:21,840 --> 00:34:25,239
or whatever, It's just you're
just trying to learn the flow

536
00:34:25,239 --> 00:34:28,440
physics.
And actually you may massively

537
00:34:28,440 --> 00:34:32,920
have way too much information by
trying to do all possible ones.

538
00:34:33,320 --> 00:34:37,159
Which then I thought, couldn't
you achieve this?

539
00:34:37,159 --> 00:34:39,760
Do you really need wings?
Couldn't you just create lots of

540
00:34:39,760 --> 00:34:42,960
random shapes with lots of
random boundary conditions?

541
00:34:43,199 --> 00:34:46,880
And as long as you take every
possible consequence of shear

542
00:34:46,880 --> 00:34:49,560
flow and pressure gradients, it
doesn't actually matter if it

543
00:34:49,560 --> 00:34:52,960
looks like a plane because
you're just trying to learn the

544
00:34:52,960 --> 00:34:56,080
physics.
That's a very good question, I

545
00:34:56,080 --> 00:34:58,200
would say.
I mean, that can be helpful

546
00:34:58,280 --> 00:35:03,480
because what you want is to be
able to explore such a latent

547
00:35:03,480 --> 00:35:07,400
representation very broadly, you
know, and that you can really

548
00:35:07,400 --> 00:35:11,160
investigate many configurations
and calculations and so on.

549
00:35:11,480 --> 00:35:15,040
At the end, well, we want to
revert back to something that we

550
00:35:15,040 --> 00:35:19,520
can understand, right?
And then probably having some

551
00:35:19,520 --> 00:35:22,520
random shapes with some random
pressure gradient and curvature

552
00:35:22,520 --> 00:35:25,880
distributions would be helpful
for the exploration because that

553
00:35:25,880 --> 00:35:28,200
would allow us to identify
different mechanisms that

554
00:35:28,200 --> 00:35:30,960
probably we haven't seen just
because they were not ideal

555
00:35:30,960 --> 00:35:33,320
dynamic, but maybe they were
helpful for other things.

556
00:35:34,240 --> 00:35:36,760
But probably in our data sets,
we want to have shapes that

557
00:35:36,760 --> 00:35:39,200
we're familiar with and
applications that we're familiar

558
00:35:39,200 --> 00:35:42,960
with just for the kind of
interpretation point of view, so

559
00:35:43,000 --> 00:35:46,760
such that we can connect it.
But from the latent space

560
00:35:46,760 --> 00:35:52,160
generation, yes, the most the
more diverse and the more crazy,

561
00:35:52,160 --> 00:35:55,280
the better.
In fact, what what we are doing,

562
00:35:55,280 --> 00:36:00,200
what we are building is systems
that the agent not only does the

563
00:36:00,200 --> 00:36:03,520
exploration but also generates
new geometries by itself.

564
00:36:04,520 --> 00:36:09,120
So we have this problem that we
like very much and we're having

565
00:36:09,120 --> 00:36:11,920
a pre print coming out very soon
on this.

566
00:36:12,160 --> 00:36:16,680
So it's basically 2 cylinders to
the keep it simple for now, but

567
00:36:16,680 --> 00:36:19,760
it's 2 cylinders with different
radii and different separations.

568
00:36:20,240 --> 00:36:25,960
And we just want the our system
agentic AI with a foundation

569
00:36:25,960 --> 00:36:32,640
model system to to explore the
wake of this tandem cylinder

570
00:36:32,640 --> 00:36:36,120
arrangement and just look at the
recovery of the wake.

571
00:36:36,640 --> 00:36:40,880
And then what this does is,
well, in order to do that, and

572
00:36:40,880 --> 00:36:42,520
this is trained on few cases,
right?

573
00:36:42,520 --> 00:36:45,560
It's not trained on many cases,
but I need to explore the whole

574
00:36:45,560 --> 00:36:47,200
space, right?
So these agents, what they do

575
00:36:47,200 --> 00:36:50,600
is, oh, but I need to look at
this configuration and calculate

576
00:36:50,600 --> 00:36:53,640
my weight characteristics and
internal quantities and you

577
00:36:53,640 --> 00:36:56,520
know, displacement thickness,
momentum thickness, okay, I know

578
00:36:56,520 --> 00:36:58,480
this.
But then I need to look at a

579
00:36:58,480 --> 00:37:00,560
different case and a different
case and a different case.

580
00:37:00,920 --> 00:37:05,080
And then it starts to
sequentially understand what the

581
00:37:05,080 --> 00:37:07,760
scaling of the wake recovery
looks like as a function of the

582
00:37:07,760 --> 00:37:11,280
parameters of your problem.
And you haven't not had to

583
00:37:11,280 --> 00:37:14,320
create hundreds of cases.
You only have to create a few

584
00:37:14,320 --> 00:37:16,920
cases.
But it's the foundation model

585
00:37:16,920 --> 00:37:20,480
combined with the agent that is
doing that autonomously is doing

586
00:37:20,480 --> 00:37:23,760
the discovery for you by
identifying what regions are

587
00:37:23,760 --> 00:37:26,960
important and creating and
generating new data in those

588
00:37:27,040 --> 00:37:28,520
regions.
In those regions.

589
00:37:29,280 --> 00:37:35,000
I think that's where we are
going so that we can even find a

590
00:37:35,000 --> 00:37:39,080
discovery of cases that we don't
even think of because, because

591
00:37:39,080 --> 00:37:43,640
we have been looking at patterns
influenced by the past history,

592
00:37:43,640 --> 00:37:45,360
right?
And kind of like the need of

593
00:37:45,360 --> 00:37:50,360
applications, but autonomously
exploring the whole space as

594
00:37:51,120 --> 00:37:53,800
agents can do.
They can really multiply the the

595
00:37:53,800 --> 00:37:57,560
the insights and new solutions
and new mechanisms that we can

596
00:37:57,560 --> 00:38:02,600
actually discover.
And and that's why I'd feel and

597
00:38:02,600 --> 00:38:06,960
I'm guilty of this myself, that
we are still probably too, too

598
00:38:06,960 --> 00:38:12,160
much splitting the fluids task
or the data generation task from

599
00:38:12,160 --> 00:38:15,840
the machine learning task.
And so, you know, I'll be

600
00:38:15,840 --> 00:38:18,920
transparent for the data sets
that I've typically generated,

601
00:38:19,560 --> 00:38:22,760
they have been kind of segmented
a little bit for practical

602
00:38:22,760 --> 00:38:26,320
reasons, for people reasons.
So you're like, right, Well, you

603
00:38:26,320 --> 00:38:28,080
know, you have a bunch of
meetings, you decide what you're

604
00:38:28,080 --> 00:38:30,040
going to do.
You then kick off the data

605
00:38:30,040 --> 00:38:33,320
generation exercise, you
generate lots of data, and then

606
00:38:33,320 --> 00:38:36,520
you train the model.
And where, as you say, what you

607
00:38:36,520 --> 00:38:39,800
really want to be doing is
constantly them being in a loop

608
00:38:39,880 --> 00:38:44,080
together and only generating the
data it needs to generate.

609
00:38:44,800 --> 00:38:47,960
But I don't know if you found
this in your work, but one of

610
00:38:47,960 --> 00:38:51,720
the problems is that even the
machine learning frameworks are

611
00:38:51,720 --> 00:38:55,600
typically not even written in
the same language that the data

612
00:38:55,600 --> 00:38:59,120
generation is.
And doing that on a big HPC

613
00:38:59,120 --> 00:39:02,360
machine and training and
translating information, it's,

614
00:39:02,360 --> 00:39:07,320
it's quite challenging and maybe
brings the topic that, you know,

615
00:39:07,320 --> 00:39:10,040
we briefly discussed before we
started recording, which is this

616
00:39:11,120 --> 00:39:16,040
fluid community, computer
science, ML community.

617
00:39:16,840 --> 00:39:21,720
How have you seen those two
communities come closer?

618
00:39:22,480 --> 00:39:25,640
Still not close enough, You
know, how how have you assessed

619
00:39:25,640 --> 00:39:27,480
this?
And, and, and to be fair, you

620
00:39:27,480 --> 00:39:31,400
have done a great deal.
And to be fair with people like

621
00:39:31,520 --> 00:39:34,720
who you collaborate, Steve
Brunton of of trying to, you

622
00:39:34,720 --> 00:39:37,200
know, bring a little bit
together.

623
00:39:38,000 --> 00:39:40,520
But yeah, how how do you see
those two communities at the

624
00:39:40,520 --> 00:39:42,480
moment?
Well, thank you.

625
00:39:43,280 --> 00:39:48,520
Thank you so much.
First, I think the problem that

626
00:39:48,520 --> 00:39:53,080
has happened for a while and
still present, although probably

627
00:39:53,080 --> 00:39:56,480
getting a bit better, is that
they have been a bit

628
00:39:57,200 --> 00:40:02,120
disconnected in really
understanding the the depths of

629
00:40:02,120 --> 00:40:06,040
the problems in fluid mechanics.
And perhaps the fluid mechanics

630
00:40:06,040 --> 00:40:09,640
community, when trying to adopt
MMM methods, has been sometimes

631
00:40:09,640 --> 00:40:14,280
a bit naive in just.
Adopting very vanilla

632
00:40:14,280 --> 00:40:17,640
architectures and maybe giving
up a bit too quickly without

633
00:40:17,640 --> 00:40:19,960
fully understanding.
So, you know, if the mechanics

634
00:40:19,960 --> 00:40:23,800
community is guilty in part, and
I think what has happened also

635
00:40:23,800 --> 00:40:28,800
from the CS community is that,
you know, I mean aerodynamics

636
00:40:28,800 --> 00:40:31,720
from I am at the aerospace
engineering department right

637
00:40:31,960 --> 00:40:34,920
here in Michigan.
We train engineers for many

638
00:40:34,920 --> 00:40:37,880
years to understand the
mechanics, aerodynamics,

639
00:40:38,280 --> 00:40:39,640
instructors and many other
things.

640
00:40:39,640 --> 00:40:41,400
But of course fluid mechanics is
a big part of it.

641
00:40:42,480 --> 00:40:46,840
So you can't understand all the
complexities of fluid mechanics,

642
00:40:48,000 --> 00:40:50,080
you know, in the course of 2-3
months, right?

643
00:40:50,080 --> 00:40:53,040
Because there is a lot to look
at.

644
00:40:53,520 --> 00:41:00,240
These are very entangled
problems and sometimes the

645
00:41:00,520 --> 00:41:04,240
questions and the problems that
we have in fluid mechanics are

646
00:41:04,240 --> 00:41:07,400
not what typically the CS
community has looked at

647
00:41:07,800 --> 00:41:11,640
necessarily.
Because you know, the type of

648
00:41:11,640 --> 00:41:15,240
data that has been used in the
CS community to develop methods

649
00:41:15,520 --> 00:41:17,720
is not necessarily of the same
characteristics as the fluid

650
00:41:17,720 --> 00:41:20,080
mechanics data.
And this is very rich data with

651
00:41:20,080 --> 00:41:23,320
a very broadband Spectra multi
scale.

652
00:41:23,520 --> 00:41:25,200
This is really, really complex
physics, right?

653
00:41:25,440 --> 00:41:31,440
So maybe looking at the MSE is
not the right metric, right?

654
00:41:31,440 --> 00:41:33,920
Because this only tells you a
very, very partial view of the

655
00:41:33,920 --> 00:41:36,120
story.
And maybe even having a model

656
00:41:36,120 --> 00:41:39,400
that predicts the MSC, well,
that might be useless in some

657
00:41:39,400 --> 00:41:41,360
cases, right?
We maybe want other things.

658
00:41:41,360 --> 00:41:44,880
Maybe we just want to look at,
you know, systems that predict

659
00:41:44,880 --> 00:41:47,560
separation, right or that
predict certain aspects of the

660
00:41:47,560 --> 00:41:52,040
mixing or maybe, you know,
certain structures, certain

661
00:41:52,040 --> 00:41:56,640
mechanisms as I was mentioning
before for, you know, for a drug

662
00:41:56,880 --> 00:42:00,760
producing mechanisms.
So it could be that for my

663
00:42:00,760 --> 00:42:04,640
purpose, if I want to reduce
drug, having a good MSE is

664
00:42:04,640 --> 00:42:07,600
actually counterproductive
because the MSE is going to

665
00:42:07,600 --> 00:42:11,560
drive me to certain events,
which may be very energetic and

666
00:42:11,560 --> 00:42:15,600
they may be contributing a lot
to the MSE, but not so relevant

667
00:42:15,600 --> 00:42:19,600
towards the drag generation,
which is the stuff that I want

668
00:42:19,600 --> 00:42:22,080
to, you know, focus on in my
application.

669
00:42:22,440 --> 00:42:27,440
Therefore, a model with a much
worse MSE but getting the error

670
00:42:27,440 --> 00:42:30,920
right in the right mechanisms is
way better, right?

671
00:42:31,160 --> 00:42:36,200
So I think the this type of this
type of interaction is what is

672
00:42:36,200 --> 00:42:39,240
really going to help us build
more useful models, at least

673
00:42:39,240 --> 00:42:40,480
from the fluid mechanics
community.

674
00:42:40,680 --> 00:42:42,560
And of course, this is not just
for fluid mechanics.

675
00:42:42,560 --> 00:42:45,240
This is applicable to any area
of, of science, right?

676
00:42:45,240 --> 00:42:50,080
But of course, I can speak, you
know, with more depth on the

677
00:42:50,080 --> 00:42:54,600
fluid mechanics problems because
what you want, and this is

678
00:42:54,600 --> 00:42:57,840
another question, you know,
let's build a foundation model.

679
00:42:57,840 --> 00:43:00,600
Let's build a big ML system for
fluids, OK.

680
00:43:00,600 --> 00:43:06,280
But for fluids for what, right.
I mean, within fluids and within

681
00:43:07,000 --> 00:43:10,880
aerodynamics and within
combustion, I mean, there are so

682
00:43:10,880 --> 00:43:14,400
many questions that we can look
at and probably the model that

683
00:43:14,400 --> 00:43:17,360
you build is going to be quite
different for all those

684
00:43:17,360 --> 00:43:20,440
questions, right?
So I think that's really

685
00:43:20,440 --> 00:43:24,800
understanding in more depth what
the fluid mechanics problems

686
00:43:24,800 --> 00:43:27,240
require.
What are the questions that we

687
00:43:27,240 --> 00:43:28,520
have?
What are the questions that we

688
00:43:28,520 --> 00:43:32,200
don't have?
How to build models that can

689
00:43:32,200 --> 00:43:36,560
really effectively tackle those
questions, that can be a much

690
00:43:36,560 --> 00:43:38,760
more effective use of
everybody's time, basically.

691
00:43:40,160 --> 00:43:47,080
Where do you sit on the debate
around a data-driven versus

692
00:43:47,080 --> 00:43:51,360
physics driven models?
You know, again, the usual

693
00:43:51,480 --> 00:43:58,400
criticism from people who are
maybe from a fluid background

694
00:43:58,400 --> 00:44:02,680
is, you know, any of these
data-driven approaches cannot

695
00:44:02,680 --> 00:44:05,640
guarantee mass conservation,
energy conservation.

696
00:44:05,640 --> 00:44:09,160
They can't guarantee, you know,
we should be implicitly or

697
00:44:09,160 --> 00:44:13,400
explicitly imposing, which I
guess is where some of the pins,

698
00:44:13,720 --> 00:44:21,080
you know, things started.
But then the successes seem to

699
00:44:21,080 --> 00:44:26,640
have been not as strong in the
sort of physics informed,

700
00:44:26,640 --> 00:44:32,040
inspired, conditioned.
So yeah, where where do you sit

701
00:44:32,040 --> 00:44:36,600
on that side of the fence?
That's, well, that's, that's a

702
00:44:36,600 --> 00:44:42,040
big question, right?
Of course you need to, you need

703
00:44:42,040 --> 00:44:45,240
to use any physical information
that you have about your system

704
00:44:45,440 --> 00:44:46,920
when building your models,
right.

705
00:44:47,480 --> 00:44:53,080
But if you just use the
equations and nothing else, then

706
00:44:53,080 --> 00:44:56,040
you're building a numerical
solver, right, with a

707
00:44:56,040 --> 00:45:01,800
data-driven numerical solver.
And if you just use data, then

708
00:45:01,800 --> 00:45:06,600
you are a problem, you know, be
being very wasteful because

709
00:45:06,600 --> 00:45:09,280
there's a lot of information
that you have about your system

710
00:45:09,600 --> 00:45:12,200
that you need to relearn.
Like, you know that your system

711
00:45:12,200 --> 00:45:15,040
has certain symmetries, certain
conservation properties.

712
00:45:15,640 --> 00:45:19,480
If you don't embed that in your
system, well, first of all, your

713
00:45:19,480 --> 00:45:22,560
mother will be wrong because you
will never be, you know,

714
00:45:22,760 --> 00:45:27,080
conservative exactly right or as
exactly as possible.

715
00:45:27,240 --> 00:45:30,400
But second, even if it's a very,
it's a system that is very

716
00:45:30,400 --> 00:45:32,880
conservative, you have to learn
that right.

717
00:45:32,880 --> 00:45:36,000
So you need to use data and
compute to get to that point.

718
00:45:36,760 --> 00:45:39,320
And of course, that's something
that you knew from the

719
00:45:39,320 --> 00:45:42,160
beginning.
So it's not very smart to to

720
00:45:42,160 --> 00:45:45,280
waste that information.
So like with everything in life,

721
00:45:45,280 --> 00:45:47,960
right, no extreme is going to be
optimal.

722
00:45:48,160 --> 00:45:50,800
And I think that both extremes
have been explored and we have

723
00:45:50,800 --> 00:45:55,080
learned a lot from it.
To me, being able to use

724
00:45:55,400 --> 00:45:58,080
symmetries, conservation laws in
your problems.

725
00:45:58,280 --> 00:46:01,840
If you're trying to solve for
several velocity components, I

726
00:46:02,000 --> 00:46:04,480
mean, do you know that there is
the flossing compressible, Do

727
00:46:04,480 --> 00:46:06,760
you know that you can use
incompressibility to get the

728
00:46:06,800 --> 00:46:10,720
velocity component or use that
in the way that you're building

729
00:46:10,720 --> 00:46:13,520
your system?
But also going back to an idea

730
00:46:13,520 --> 00:46:17,200
that I like very much, this
whole explainability, causality

731
00:46:17,200 --> 00:46:21,200
mechanisms.
If we know these things, why

732
00:46:21,200 --> 00:46:24,520
don't we try to build systems
that are really focused on this?

733
00:46:25,240 --> 00:46:29,200
And one example is, let's
imagine that I want to create a

734
00:46:29,200 --> 00:46:33,800
system that predicts very well
the acoustic feel of an airfoil

735
00:46:33,800 --> 00:46:35,760
or a drone, right?
Because drones can be very

736
00:46:35,760 --> 00:46:37,600
noisy.
I want to be able to control the

737
00:46:37,600 --> 00:46:40,600
acoustic feel, the noise
produced by this drone.

738
00:46:41,360 --> 00:46:46,280
Why don't I use some of these
causality explainability methods

739
00:46:46,280 --> 00:46:50,520
to identify and pinpoint the
mechanisms producing the noise?

740
00:46:50,520 --> 00:46:53,320
And by mechanisms, I'm talking
about flow regions interacting

741
00:46:53,320 --> 00:46:55,320
with each other in physical
space.

742
00:46:55,920 --> 00:46:59,760
And then when I build a
predictive model, I don't look

743
00:46:59,840 --> 00:47:04,640
at the, you know, trying to get
neither just the governing

744
00:47:04,640 --> 00:47:08,080
equations like in a pinch way or
just use data to predict the

745
00:47:08,080 --> 00:47:10,400
flow fields.
But rather, why don't I try to

746
00:47:10,400 --> 00:47:13,640
predict just those extractors,
those mechanisms that produce

747
00:47:13,640 --> 00:47:18,800
the acoustic field in that way,
you're really encapsulating in

748
00:47:18,800 --> 00:47:21,680
those structures a lot of
information.

749
00:47:22,000 --> 00:47:24,480
And that's where I think that
one can have computational

750
00:47:24,480 --> 00:47:27,520
savings with respect to, you
know, doing a much more resolved

751
00:47:27,520 --> 00:47:30,120
simulation.
Because those structures

752
00:47:30,120 --> 00:47:33,760
producing the acoustic field are
the result of a lot of, you

753
00:47:33,760 --> 00:47:36,240
know, interscale energy
interactions that you need to

754
00:47:36,240 --> 00:47:42,880
resolve when you do ADNS.
But the model, I mean, once that

755
00:47:42,880 --> 00:47:46,760
DNS is run, those structures
that are producing the acoustic

756
00:47:46,760 --> 00:47:49,640
field and so on, that the result
of those interactions, right?

757
00:47:49,640 --> 00:47:52,520
So you don't need to re simulate
them all the time.

758
00:47:52,760 --> 00:47:55,840
Once you know what the
mechanisms are, you can try to

759
00:47:55,840 --> 00:47:59,960
develop models that target those
physical mechanisms precisely

760
00:48:00,080 --> 00:48:04,000
and not everything else.
Of course.

761
00:48:04,000 --> 00:48:07,720
That's why I repeat, sometimes I
want a model for what?

762
00:48:07,800 --> 00:48:10,440
If your model is for the
acoustic field, then my advice

763
00:48:10,440 --> 00:48:12,800
would be, well, identify what
are the mechanisms producing the

764
00:48:12,800 --> 00:48:15,560
acoustic field and try to get
those very well, right?

765
00:48:15,560 --> 00:48:18,760
Because probably you will not
get a perfect model, but you

766
00:48:18,760 --> 00:48:21,520
will get something that is
pretty accurate and pretty

767
00:48:21,760 --> 00:48:23,600
efficient because you don't need
to solve everything.

768
00:48:24,280 --> 00:48:27,400
But if your goal is to have
something that gets the acoustic

769
00:48:27,400 --> 00:48:30,480
field very well and the gusts
and the turbulence and the heat

770
00:48:30,480 --> 00:48:34,600
transfer well, then you need a
multi physics DNS, right?

771
00:48:34,600 --> 00:48:36,120
Then, then you need to solve
everything.

772
00:48:36,400 --> 00:48:40,040
So it depends a little bit on
what you want and of course,

773
00:48:40,600 --> 00:48:45,240
embedding that physical insight
in your model, not as equations,

774
00:48:45,240 --> 00:48:50,280
but as mechanisms that you can
then, you know, recreate with

775
00:48:50,280 --> 00:48:52,720
your system.
I think that that's a pretty

776
00:48:52,720 --> 00:48:58,080
promising way to go.
And how from a, you know,

777
00:48:58,080 --> 00:49:00,520
you're, you're obviously a
professor at university.

778
00:49:01,560 --> 00:49:04,240
What do you think?
What's the role of academia in

779
00:49:04,240 --> 00:49:11,160
this, focusing more on like the
students in the education side,

780
00:49:11,240 --> 00:49:15,320
you know, do you try and team up
more with the computer science

781
00:49:15,320 --> 00:49:19,800
departments?
Is it a matter of swapping

782
00:49:19,800 --> 00:49:23,200
people over like, but but then
equally, I guess you don't want

783
00:49:23,200 --> 00:49:26,520
to not teach some of your
existing content.

784
00:49:26,520 --> 00:49:29,760
So how do you balance getting
rid of some content, bringing

785
00:49:29,760 --> 00:49:35,040
new ones in?
That's that's a key question.

786
00:49:35,360 --> 00:49:37,320
And we're having many
discussions around this.

787
00:49:38,480 --> 00:49:41,160
For example, here in Michigan, I
mean, we're having a lot of

788
00:49:41,160 --> 00:49:46,160
conversations around our new
curriculum and how to well adapt

789
00:49:46,160 --> 00:49:51,920
to AI era, LLM base era.
So we're having many

790
00:49:51,920 --> 00:49:55,840
conversations around this.
So first of all, we have a very

791
00:49:55,840 --> 00:49:59,360
good interaction with Los Alamos
National Lab.

792
00:49:59,360 --> 00:50:05,040
We actually have a bunch of Los
Alamos scientists sitting on our

793
00:50:05,040 --> 00:50:08,280
campus and working with them
daily and interacting with them

794
00:50:08,440 --> 00:50:11,280
constantly.
So we really get access to that

795
00:50:11,280 --> 00:50:15,720
synergy and that cooperation
with computer scientists, which

796
00:50:15,720 --> 00:50:20,920
are world class, right?
So with with really incredible

797
00:50:20,920 --> 00:50:24,000
computational facilities.
So I think such an interface and

798
00:50:24,000 --> 00:50:26,600
such a close connection is, is
essential.

799
00:50:26,880 --> 00:50:31,560
But the other thing is from the
educational point of view, we

800
00:50:31,560 --> 00:50:35,480
need to be of course aware of
the fact that students have

801
00:50:35,480 --> 00:50:38,800
access to LLMS, right?
And and that, you know, the way

802
00:50:38,800 --> 00:50:42,440
that we teach, the way that we
assess is different from what it

803
00:50:42,440 --> 00:50:45,920
used to be.
But in a way, and this is

804
00:50:46,400 --> 00:50:49,000
conversations that we're having
with my colleagues and I think

805
00:50:49,000 --> 00:50:51,640
that there's quite some
interesting, I'm promising

806
00:50:51,640 --> 00:50:56,440
directions in here.
We, we don't need to necessarily

807
00:50:56,440 --> 00:51:00,280
see this as a disadvantage, but
rather as an opportunity to use

808
00:51:00,280 --> 00:51:05,320
these systems to encourage
students for more exploration,

809
00:51:05,320 --> 00:51:12,000
for deeper, for, for deeper
questions and deeper assessments

810
00:51:12,000 --> 00:51:15,240
of the content almost in a
customized way, right?

811
00:51:15,240 --> 00:51:19,280
So, so I think that rather than
thinking that students are going

812
00:51:19,280 --> 00:51:22,360
to be lazy and just look up
their answers and not think,

813
00:51:22,720 --> 00:51:26,120
rather use these these systems
to help students think in a way.

814
00:51:28,480 --> 00:51:32,440
But do you feel that?
I guess flicking around the

815
00:51:32,440 --> 00:51:36,560
other way, do do computer
science students need to learn

816
00:51:36,560 --> 00:51:39,280
more about the sciences?
You know, there's a lot of like

817
00:51:39,440 --> 00:51:42,240
AI for science.
If the engineers are all

818
00:51:42,240 --> 00:51:44,880
learning some of the computer
science being, then what are the

819
00:51:44,880 --> 00:51:48,360
computer scientists doing?
Is there, is there an equal

820
00:51:48,640 --> 00:51:50,760
thing on the computer science
say hey you all need to be

821
00:51:50,760 --> 00:51:53,360
learning some science domain
because AI is going to do the

822
00:51:53,360 --> 00:51:55,800
some of the core CS stuff, you
know?

823
00:51:56,680 --> 00:52:03,160
I was, well, I'm, I can tell you
that there is, I mean, there is

824
00:52:03,160 --> 00:52:07,600
an increase of enrollment of
students in, in areas like

825
00:52:07,600 --> 00:52:12,200
aerospace engineering and, and
part of it comes from computer

826
00:52:12,200 --> 00:52:14,200
science, although of course the
computer science degrees,

827
00:52:14,480 --> 00:52:18,600
they've been excellent at
adapting to, to new technologies

828
00:52:18,600 --> 00:52:22,400
and new well, developments
within AI for science.

829
00:52:23,400 --> 00:52:27,880
But there is more and more
overlap and more and more, you

830
00:52:27,880 --> 00:52:32,400
know, joint educational programs
across systems and across

831
00:52:32,720 --> 00:52:36,160
engineering disciplines.
And yes, I would, I would agree

832
00:52:36,160 --> 00:52:40,080
completely that if computer
scientists are going to be the

833
00:52:40,520 --> 00:52:45,920
developing AI for science or AI
for engineering methods, then

834
00:52:45,920 --> 00:52:48,760
they're going to need more
background in these, in these

835
00:52:48,760 --> 00:52:50,560
topics.
And that's something that is

836
00:52:50,560 --> 00:52:57,720
really that's multidisciplinary
wealth and and richness is

837
00:52:57,720 --> 00:53:02,440
what's going to bring really
stellar research in the next

838
00:53:02,440 --> 00:53:04,560
years, right?
That 1 is really capable of

839
00:53:04,560 --> 00:53:07,800
finding completely new
directions like the foundation

840
00:53:07,800 --> 00:53:11,120
model and the agent are finding
patterns across data sets that

841
00:53:11,120 --> 00:53:14,720
you would not imagine, right.
When we start to interact in

842
00:53:14,720 --> 00:53:17,800
disciplines at a different
level, we're going to find such

843
00:53:18,320 --> 00:53:20,520
connections that maybe we did
not expect at the beginning.

844
00:53:22,360 --> 00:53:25,640
So one of the things that, you
know, you mentioned a few times

845
00:53:25,640 --> 00:53:29,720
is on the like control and
optimization side of things.

846
00:53:30,120 --> 00:53:33,760
And I feel at least myself a
little bit guilty that I, I'm

847
00:53:33,760 --> 00:53:37,080
always thinking of, you know,
the, the impact of machine

848
00:53:37,080 --> 00:53:39,880
learning or reduced order models
in terms of a predicted

849
00:53:39,880 --> 00:53:41,920
capability.
You know, we can predict this

850
00:53:41,920 --> 00:53:47,160
flow much faster.
But I guess for industry or even

851
00:53:47,160 --> 00:53:49,880
for scientific discovery, but
particularly for industry, it's

852
00:53:49,880 --> 00:53:53,680
almost always a kind of
optimization problem in a way

853
00:53:53,680 --> 00:53:57,720
isn't it's like I need to have
the best design or the lowest

854
00:53:57,720 --> 00:54:01,160
drag or the it's never just a
prediction.

855
00:54:02,360 --> 00:54:06,640
So, So what progress have you
seen with the sort of automatic

856
00:54:06,640 --> 00:54:10,280
differentiation, the the ability
of machine learning to maybe

857
00:54:10,280 --> 00:54:14,200
help advance that optimization
process?

858
00:54:15,000 --> 00:54:18,760
Yeah, I I think that prediction
is important.

859
00:54:18,760 --> 00:54:20,960
It's interesting.
It's not the most impactful

860
00:54:20,960 --> 00:54:24,360
application of machine learning.
It's in optimization and

861
00:54:24,360 --> 00:54:26,200
control.
That's where we can really have

862
00:54:26,760 --> 00:54:28,320
the, the biggest, the biggest
impact.

863
00:54:28,600 --> 00:54:31,320
So as I mentioned before, I, I
worked a lot on reinforcement

864
00:54:31,320 --> 00:54:36,040
learning.
So that's an area where we have

865
00:54:36,040 --> 00:54:41,400
been able to control cases that
were not possible three or four

866
00:54:41,400 --> 00:54:45,720
years ago or at least with this
degree of control authority.

867
00:54:47,840 --> 00:54:50,840
And not only.
So I would say not just from a

868
00:54:50,840 --> 00:54:55,960
purely design point of view of
reducing drag or enhancing

869
00:54:55,960 --> 00:54:59,120
mixing.
I would argue that for

870
00:54:59,600 --> 00:55:03,280
discovery, for really
understanding the mechanisms and

871
00:55:03,280 --> 00:55:07,520
the building blocks of these
physical systems, you can use

872
00:55:07,520 --> 00:55:09,960
optimization and, and, and
reinforcement only.

873
00:55:09,960 --> 00:55:16,000
For example, I mean, imagine a
tumulant flow where a made-up

874
00:55:16,000 --> 00:55:21,800
turbulent flow where the energy
fluxes are restricted to certain

875
00:55:22,480 --> 00:55:26,240
scales and to certain work
numbers dynamically, right That

876
00:55:26,240 --> 00:55:30,600
you do it as the flow is running
and, and, and you want to do

877
00:55:30,600 --> 00:55:32,920
that in physical space.
So you want to really affect

878
00:55:32,920 --> 00:55:35,040
certain instructors with a
certain forcing.

879
00:55:35,400 --> 00:55:38,120
Well, you are probably going to
have to do it with some

880
00:55:38,120 --> 00:55:40,280
optimization technique.
And reinforcement learning is

881
00:55:40,280 --> 00:55:42,840
good for optimizing systems that
are changing dynamically.

882
00:55:43,200 --> 00:55:47,160
So you you're going to end up
with a with a modified system or

883
00:55:47,160 --> 00:55:50,800
a system where production is
minimized and dissipation is

884
00:55:50,800 --> 00:55:53,360
maximized.
How does a flow like that look

885
00:55:53,360 --> 00:55:56,000
like?
And let me study, let me study

886
00:55:56,000 --> 00:55:58,040
that new flow, right?
I mean, I have created a

887
00:55:58,040 --> 00:56:00,880
completely made-up flow thanks
to optimization.

888
00:56:01,640 --> 00:56:03,720
And now I can look at the
Spectra, I can look at the

889
00:56:03,720 --> 00:56:06,840
structures and learn something
about my problem, right.

890
00:56:07,160 --> 00:56:12,720
So yes, these capabilities with
optimization are key also for

891
00:56:12,720 --> 00:56:14,480
discovery and so for getting
insight.

892
00:56:14,800 --> 00:56:18,040
And going back to your previous
point about how new systems and

893
00:56:18,040 --> 00:56:22,760
differential solvers, I mean,
we, we are really having the

894
00:56:22,760 --> 00:56:30,120
chance of solving problems at
scale of this optimization type

895
00:56:30,120 --> 00:56:32,680
of system, which was not
possible before.

896
00:56:33,040 --> 00:56:37,920
And inverse problems where we
can, you know, optimal sensing,

897
00:56:37,920 --> 00:56:41,200
we can really look at optimal
initial conditions and optimal

898
00:56:41,200 --> 00:56:45,280
perturbations really for very,
very challenging configurations.

899
00:56:45,920 --> 00:56:49,600
All of that is possible now
thanks to the scale that we can

900
00:56:49,600 --> 00:56:54,200
achieve with these systems.
So I think that's, yeah, I mean,

901
00:56:54,200 --> 00:56:56,280
prediction is, is good.
It's interesting.

902
00:56:56,440 --> 00:57:02,600
It's only part of the problem.
It's really in the control and

903
00:57:02,600 --> 00:57:06,080
optimization and in the
explainability of those

904
00:57:06,080 --> 00:57:09,360
predictions.
Like, OK, I have predicted very

905
00:57:09,360 --> 00:57:12,080
well in my system, but why are
those predictions happening?

906
00:57:12,080 --> 00:57:15,520
And what are the really
important physics and effects

907
00:57:15,720 --> 00:57:16,880
that I can observe with my
system?

908
00:57:17,160 --> 00:57:19,320
That's that's the stuff that I
think is is is actually

909
00:57:19,320 --> 00:57:22,760
valuable.
So maybe a sort of final

910
00:57:22,760 --> 00:57:27,040
question or, or topic looking
towards the future a little bit.

911
00:57:27,040 --> 00:57:30,000
You, you said right at the
beginning that if, if I had

912
00:57:30,000 --> 00:57:33,720
asked you a few years ago, you'd
have said that we're, we're

913
00:57:33,720 --> 00:57:35,560
actually, you know, further
away.

914
00:57:35,960 --> 00:57:44,360
So if you look forward in time
and let's say to 2030, so 4

915
00:57:44,360 --> 00:57:49,920
years out, yeah, four years out,
where do you think we'll be at?

916
00:57:49,920 --> 00:57:54,840
Do you perceive there being a a
sort of ChatGPT moment in fluid

917
00:57:54,840 --> 00:57:59,160
or do you think it'll be just
more incremental improvements?

918
00:57:59,480 --> 00:58:06,680
Yeah, that's a tough question.
I think one key to go towards

919
00:58:06,680 --> 00:58:12,000
ChatGPT moment in fluids was to
realise that we need a good

920
00:58:12,000 --> 00:58:15,640
latent representations for our
problem.

921
00:58:15,760 --> 00:58:18,000
And I think that now there's
several groups of several people

922
00:58:18,000 --> 00:58:20,760
working in that direction.
So I think we're probably in the

923
00:58:20,760 --> 00:58:23,600
right path.
I would not.

924
00:58:25,000 --> 00:58:30,400
So I don't know if we want to
aim at being having a ChatGPT

925
00:58:30,400 --> 00:58:34,520
situation in fluids, maybe
something beyond maybe something

926
00:58:35,000 --> 00:58:37,800
more than what chat PT can can
do.

927
00:58:39,640 --> 00:58:43,120
What I think that we are aiming
at, and I don't know if that

928
00:58:43,120 --> 00:58:46,760
would be in 20-30, but I think
that it's actually reasonable to

929
00:58:46,760 --> 00:58:50,120
think that it should happen in
the next years, is more

930
00:58:50,120 --> 00:58:52,600
autonomous discovery, a more
automatic discovery.

931
00:58:53,440 --> 00:59:01,240
So one, I like the, the one
example of how big changes in in

932
00:59:01,240 --> 00:59:04,240
the paradigm of physics have
taken place and of course, well,

933
00:59:04,520 --> 00:59:07,560
relativity departing from more
classical mechanics.

934
00:59:07,920 --> 00:59:13,000
I mean, it takes, if you think
about it, a lot of a lot of

935
00:59:13,200 --> 00:59:17,240
coincidences to line up in order
for that to happen.

936
00:59:17,600 --> 00:59:21,920
And you need someone who is
smart enough and also crazy

937
00:59:21,920 --> 00:59:24,080
enough to propose something very
different.

938
00:59:24,640 --> 00:59:27,440
And of course, Einstein lacked
all the math background to be

939
00:59:27,440 --> 00:59:32,000
able to, you know, develop the
theory of relativity properly.

940
00:59:32,000 --> 00:59:34,560
So he had to learn the math and
interact with the right people

941
00:59:34,560 --> 00:59:38,120
to be able to do this properly.
And then, you know, to have the

942
00:59:38,120 --> 00:59:42,320
right experiments of the of the
solar eclipse to really look at

943
00:59:42,320 --> 00:59:46,240
the deviation of the sand beams.
And there's some fun stories

944
00:59:46,240 --> 00:59:49,320
about how those measurements
took place because there were

945
00:59:49,480 --> 00:59:51,880
several groups doing the
experiments at the same time,

946
00:59:52,200 --> 00:59:55,200
and they had different types of
problems to really get the

947
00:59:55,200 --> 00:59:59,120
measurements that eventually
corroborated that light was, you

948
00:59:59,120 --> 01:00:03,360
know, displaced by the sun in a
way consistent with relativity.

949
01:00:03,720 --> 01:00:09,240
There's a lot of certain Dipion
discovery, a lot of coincidences

950
01:00:09,240 --> 01:00:14,360
lined up, and a lot of human try
and error to achieve truly

951
01:00:14,360 --> 01:00:16,320
transformative discovery in
science.

952
01:00:16,840 --> 01:00:20,960
And I think that the sort of
systems where we can line up on

953
01:00:20,960 --> 01:00:24,640
the same space, the data from
different disciplines and having

954
01:00:24,640 --> 01:00:30,360
agents systematically probing
and analysing these data sets,

955
01:00:30,840 --> 01:00:34,440
that can really be the way in
which we can accelerate that.

956
01:00:34,440 --> 01:00:37,640
But it's this acceleration of
discovery, which is something

957
01:00:37,640 --> 01:00:40,760
that, you know, many people are
talking about with different

958
01:00:40,760 --> 01:00:42,680
meanings.
And I think it's a little bit

959
01:00:42,680 --> 01:00:45,400
shallow the way that it's been
portrayed sometimes.

960
01:00:45,960 --> 01:00:52,600
To me, what it means is being
able to probe the data and find

961
01:00:52,600 --> 01:00:58,600
patterns in a more systematic
way than what it takes us as

962
01:00:58,600 --> 01:01:01,720
humans to take those leaps.
Because we're it's really

963
01:01:01,720 --> 01:01:06,280
decades and decades and
centuries to line up all those

964
01:01:06,280 --> 01:01:09,600
coincidences necessary for that
leap to happen.

965
01:01:10,280 --> 01:01:14,200
And if we can do this in a way
that is much more autonomous and

966
01:01:14,200 --> 01:01:18,800
much more systematic in that
sense, that can be accelerated,

967
01:01:18,800 --> 01:01:23,040
right?
And this sort of systems I, you

968
01:01:23,040 --> 01:01:26,920
know, I think that we are really
getting there actually.

969
01:01:27,080 --> 01:01:32,440
And, and, and this is not really
about replacing human insight or

970
01:01:32,440 --> 01:01:34,360
human input.
This is really about

971
01:01:34,360 --> 01:01:37,680
accelerating the exploration on
the capability of coming up with

972
01:01:37,680 --> 01:01:42,680
mechanisms with hypothesis with
a really patterns that are non

973
01:01:42,680 --> 01:01:45,760
trivial across across
disciplines and across cases.

974
01:01:46,320 --> 01:01:50,040
So yeah, I think that we're
heading in that direction of

975
01:01:50,040 --> 01:01:55,160
really being able to make
autonomous discovery thanks to

976
01:01:55,720 --> 01:01:58,840
the possibility of compressing
data and interrogating it

977
01:01:58,840 --> 01:02:06,400
systematically in a way.
Yeah, it does seem that there is

978
01:02:06,400 --> 01:02:10,600
this, I would say inflection
point now where I would largely

979
01:02:10,600 --> 01:02:12,040
agree with you, you know, a few
years.

980
01:02:12,040 --> 01:02:14,120
Ago when I.
Started to think about some of

981
01:02:14,120 --> 01:02:17,760
like a foundational model, you
know, it it did for various

982
01:02:17,760 --> 01:02:22,920
reasons just felt quite Yeah,
just like impossible almost.

983
01:02:22,920 --> 01:02:27,640
It would, you know, may still
be, but I I also get the feeling

984
01:02:27,640 --> 01:02:34,200
that there is a, a general
movement of replicating the sort

985
01:02:34,200 --> 01:02:37,000
of chat chi BT success, you
know, the the set.

986
01:02:37,000 --> 01:02:39,760
And this is broader than just
fluids across science, across

987
01:02:39,760 --> 01:02:45,120
physics, across all, you know,
possible disciplines that, you

988
01:02:45,120 --> 01:02:48,960
know, it worked for that which
and, and I'm not I haven't

989
01:02:48,960 --> 01:02:51,680
studied a a great deal, but I,
you know, I know enough to know

990
01:02:51,680 --> 01:02:55,240
that many people thought that it
couldn't be possible, right.

991
01:02:55,600 --> 01:02:57,960
You know, people weren't saying,
oh, chat chi BT Oh yeah, we knew

992
01:02:57,960 --> 01:03:00,280
that was going to come.
You know, it was quite a big

993
01:03:00,800 --> 01:03:05,280
impact and, and it, it was done
at a scale that was larger than

994
01:03:05,280 --> 01:03:09,640
anyone thought possible.
And I, I do have a feel that

995
01:03:10,280 --> 01:03:13,560
made people more ambitious, you
know, and, and think, well, if,

996
01:03:13,800 --> 01:03:16,000
if it could work for that, it,
it could work.

997
01:03:16,440 --> 01:03:21,280
But I do like your argument on
that.

998
01:03:21,280 --> 01:03:24,400
Sometimes you need to look at
things a bit differently because

999
01:03:24,400 --> 01:03:29,560
I have a funny feeling that the
approach that ultimately makes

1000
01:03:29,560 --> 01:03:34,440
it may not be the one that is
mainstream today.

1001
01:03:34,640 --> 01:03:39,560
You know that that like in all
in science, there's always, it's

1002
01:03:39,560 --> 01:03:42,400
always the person who takes a
little bit of a, an odd

1003
01:03:42,400 --> 01:03:46,600
direction at the time which
turns out to actually be the

1004
01:03:46,600 --> 01:03:51,600
one, you know?
An agent to come up with that

1005
01:03:51,600 --> 01:03:53,360
direction, maybe.
But that's what I meant.

1006
01:03:53,360 --> 01:03:59,360
Like the, I can totally see how
with humans sort of in the loop,

1007
01:03:59,360 --> 01:04:04,760
you know, in, in, in terms of
guiding things, that the maybe

1008
01:04:04,760 --> 01:04:09,720
it is ultimately the power of
ChatGPT to find the next

1009
01:04:09,720 --> 01:04:12,920
ChatGPT.
You know, that those LLM

1010
01:04:12,920 --> 01:04:15,760
technology and the agents and
the way they work together with,

1011
01:04:15,760 --> 01:04:18,840
of course, the way that you
structure the data, you know,

1012
01:04:18,960 --> 01:04:20,720
may help.
And, and then maybe that's

1013
01:04:20,880 --> 01:04:23,160
artificial general intelligence.
And then we can all just, you

1014
01:04:23,160 --> 01:04:26,680
know, go and retire and, and,
and chill out.

1015
01:04:28,840 --> 01:04:33,440
But yeah, I really appreciate
having the chance to, to, to

1016
01:04:33,440 --> 01:04:37,400
chat about this.
What I'm going to do for people

1017
01:04:37,400 --> 01:04:42,400
listening or watching is to put
a bunch of links into the, the

1018
01:04:42,400 --> 01:04:46,800
comments because there's a few
papers that you alluded to that

1019
01:04:46,800 --> 01:04:50,120
are really good, you know, great
reads that talk in more detail

1020
01:04:50,440 --> 01:04:53,880
about some of that explainable
AI and some of the work that,

1021
01:04:53,880 --> 01:04:56,240
you know, you've done.
So I'll, we'll put them in the

1022
01:04:56,240 --> 01:05:01,120
link and highly recommend that
people read through them to go

1023
01:05:01,120 --> 01:05:04,120
even deeper than we discussed
now.

1024
01:05:04,400 --> 01:05:06,600
But yeah, I, I really appreciate
it.

1025
01:05:06,600 --> 01:05:10,280
And I think your work is going
to be seen as a major

1026
01:05:10,280 --> 01:05:14,280
contributing factor to hopefully
achieving that ChatGPT mode.

1027
01:05:14,480 --> 01:05:15,960
Well, thank you very much.
I appreciate it.

1028
01:05:15,960 --> 01:05:20,640
Maybe that's a bit optimistic,
the money work, but I appreciate

1029
01:05:20,640 --> 01:05:24,560
that much.
And at the end we, we have fun

1030
01:05:24,560 --> 01:05:26,200
with what we do.
I think that's kind of like the

1031
01:05:26,200 --> 01:05:28,280
key, the key idea and we keep
learning.

1032
01:05:28,360 --> 01:05:31,920
So that's that's pretty much it.
Exactly, exactly.

1033
01:05:32,480 --> 01:05:33,760
Great.
Thank you so much.

1034
01:05:34,120 --> 01:05:34,640
Thank you so much.