Are AI Agents and Foundation Models About to Rewrite CAE?

0:00 Hi, and welcome to the Neil Ashton Podcast. In each episode, we explained some of the fascinating ways that science and engineering are changing the world around us. We talked to leading engineers from elite level sports like cycling and Formula One to some of the world's top academics to understand how fluid dynamics, machine learning, supercomputing are bringing in a new era of discovery. We also hear some of their life stories, their career advice, the lessons they've learned on the way that I hope will be helpful to you too. So sit back and enjoy this episode. Hi and welcome back to the Neil Ashton Podcast. It's been a while, but I'm

0:43 excited to start a new season and series and I'm going to start with a discussion on how I think agents and foundation models are are really a disruptor to the current CAE and EDA ecosystem. This, this is of course is just a personal opinion. This does not represent my current employer, NVIDIA, but I wanted to share some, some thoughts I've really had over the past six months of how I see this technology being, you know, obviously a massive opportunity, but also also a disruptor. And so we'll, we'll walk through that today from a technical point of view, from a commercial point of view, from a scientific point of view.

1:26 And I'd certainly be interested to see whether you agree with my, my thinking on this. If we take a step back and you know, I'm going to look at as often I do in this, in this podcast through a sort of CFD lens, but I think much of what I'm saying is relevant across CAE and and also EDA for chip design, etcetera. If you look, where did things begin? I would argue in the 60s and the 70s people were writing their own codes. Now, I've had episodes with Professor Jameson and others who who started writing codes themselves before modern supercomputing was around, and those often started at academic institutions. Those codes then obviously

2:16 started to become useful and there was collaborations with industry, particularly in the aerospace initially. And that is why that then those aerospace companies started to work closely with those developers and eventually started to take them as their own codes. And that's why you see today even there's a legacy of aerospace companies having their own codes written from scratch, frankly, because there were no other options. And that was the way to do it. The automotive sector then took some things from the aerospace sector. And this depends on this a little bit, whether we're talking about structural mechanics or fluid mechanics,

2:55 but it started off being you had to write it yourself with some expert developer. Then over the course of, you know, the 70s, eighties and early 90s, those codes started to become commercialized. Those professors out of places like Imperial or MIT or Stanford started to, to, to create companies. And, and ultimately those are the, the lineage, the lineage of where Abacus comes from, where Fluent comes from, where Openfront comes from. And in some ways those commercialization was because it was a natural thing as one computing started to become and make these tools more than just a, a niche application. HPC facilities, you know, Beowulf cluster decided to

3:46 become commonplace and and of course, industry started to see the benefit of those simulation approaches. And frankly, then there was also a competitive nature that once one company starts, you start having competitions between, you know, the CD adapt Cos and the, and what what is now the, the Ansys and the later the Dassaults. There was that growing ecosystem of companies that that compete against each other, that accelerates it. The computing starts to get better and other than those main players, you also started to have the hardware companies get involved. So the Intel's, the AMD, the Dell, the Lenovo, the Cray in those days, the the main people

4:28 building those those systems, those supercomputing systems. And if you look at how it happened, that was probably true for, for maybe 5 to 10 years or so where it was mainly a creating software to run on workstations, servers and supercomputers. And that, you know, spread across the world, frankly. And those tools were very widely used and started to create a real, you know, hundreds of millions or even billions by then industry. The the next technology that really came around was then around like cloud computing. So cloud computing maybe 5-6, seven years ago was, was a disruptor because it allowed and, and obviously I'm skipping huge other things that happened,

5:17 but I don't want to spend 2 hours just giving the history of, of CFD. The reason I pick out cloud, and I saw it first time working for AWS, is it changed it from being a purely software license point of view to where these providers could also have a control of the hardware. Packaging software and hardware together is attractive for a company for an ISV because they can make profit off the hardware and they can lock it together both from a performance point of view, but also from a sort of stickiness point of view. And there's advantages for the customer in terms of making it more seamless because often if you just ship a binary, they

6:02 have to be the one to then figuring out how to get the hardware. They are the ones who've got to think about, do they need to buy twice as much hardware? If they want to run twice as many simulations in the theory of the ultimate SAS, then it is fully scalable. So if you're running, you know, star CCM or fluent or power flow or whatever, you could essentially say, I want to run 10 times more and through the, you know, infinite scale or inferior of the cloud you can do in the back end. So that sassification was, was something started maybe 6-7 years ago. And, and obviously companies like Rescale were one of the first to do that at a cross

6:44 software level. And there's others like Total, CAE, etcetera who tried to, to bring it together. That is where I would say AWS and Microsoft and Google started to enter into more seriously the CAE and EDA market as important partners. Because frankly, now it wasn't just The Dells and Lenovos, but you were deploying it on AWS or GCP or or Azure or or Oracle, for example. Now GPUs obviously had been around for quite a few years before that, but but arguably it's only been in the past sort of three, maybe four years where they have risen to become a really important part of the ecosystem. And frankly, that's because a lot of people have seen quite

7:28 clear benefits and they're moving to the GPUs. And that's another technology point that's important because it added a new competitive angle to this because they needed to able to run things on a GPU, which meant that if you could optimize your code for GPU can get advantage. This is where flex compute, this is where luminary cloud started to come out. The rise of, you know, VC funding for these sort of companies led to a resurgence of competitiveness, I would say, in that market. And I the reason I point to the cloud before is that many of those companies lent on as the cloud being an additional sort of angle to their to their offering.

8:12 So at that point another thing happens which is that the bigger companies start to look to consolidate and to start their sort of acquisition period. So you saw companies like Siemens and the timing of these are not all correct, but you know Siemens and Altair, Ansys and Synopsys, Cadence and Beta CAE and point wise are future facilities. Many of these companies started to acquire each other. And why was that done? Well actually so those companies then wanted to start quiet and of course this is where one of the key technology came in that starts to get to my point here,

9:00 which is the AI side. It was, it has been around to be fair, for maybe six years or even 7 years. But really it was since the sort of ChatGPT moment in the early twenty 20s, 2021-2022 that brought this to life. And so, you know, they started some new startups to really get hold and the newer concept, the Physics X, the Navisto Extraity MEAI, these companies started to come about offering the surrogate modelling capability, the ability to do real time simulations. And that frankly was a threat and an opportunity. And I say that because that is what I've heard when in conversations.

9:50 It's a threat because anything new is by definition a potential threat. These software companies who have, you know, obviously had quite a, a good position, see as you know, how do we deal with this? But clearly it's also a massive opportunity because it brings new technology, new capabilities to, to, to customers. And that it has been a very interesting thing to observe in the past, let's say 3 or 4 years of how do these larger companies deal with it? How do companies deal with it who have their own codes? How do they respond? Do they respond with hype? Do they respond with negativity, with positivity? And what happened was then some

10:29 companies decided to start developing themselves, some started to buy. So Ansys bought Extrality, which is now SIM AI. Autodesk bought Navistone, Siemens, you could argue and some part of it was the Alta acquisition. And then some of these start off start-ups that started with a more GPU focus started to pivot quite heavily to the surrogate modelling. So Lumia Cloud or Lumia AI now and Flex compute, both of you know pivoted to, to, to that direction. And so we, so we're at that point now where the, the ecosystem is at a very influential point for a sort of inflection point. And there's one final bit that

11:17 is important to that. And it's actually ultimately the main topic that I want to speak about today, which is agents. Because the missing piece to some of this has been if you create a surrogate model and you take it to its fullest extent and you have some sort of foundation model, you haven't really got rid of the need to do CAD. You haven't got rid of the need to do some sort of preprocessing or post processing. You could argue you've just got a faster plug in replacement for the for the bit of the solver or the solve part, but you still need all the bit around the rest. And that's why it's ultimately been more of an integration play through some of these AI

12:06 start-ups because you're not going to replace CFD or FEA or Eva, you're just really creating an additional tool. So in that sense you can understand why the larger Isvs and others have not necessarily seen this surrogate modelling as a massive threat initially because they know that on it's own it's it's not a replacement that has led to of course 1 angle, which is if you were to make a model that was so generalizable, then it could given the right plug and play, disrupt some solver only companies. If you only produce a solver and not the CAD and maybe not the preprocessing, then arguably you could, you know, replace some of

12:54 it. But the technology that I think it adds a real new dimension to this is the agentic side. Because what are what are agents? Well, first of all, let's tie it back to the usefulness of Chachi PT or Gemini. We have all found them to be incredibly useful in our day-to-day tasks, but I have personally noticed a a huge shift in usefulness with the agentic side. And what do we mean by the agentic side? At least the way I define it. Well, I think of it is just simply now the LLM is able to call some tools. Now that's probably oversimplifying it, but from a practical point of view, instead of me just being able to ask some ask Gemini to do something

13:47 for me or, or codecs or whatever. Although the use of codecs now is sort of murky in the waters a little bit. If I stick to like a chat prompt, it's only going to be able to really do things that rely just on text or to A to a limit where what I really want it to do is to be able to go and do things for me. So for example, if I ask, I want you to create a summary of all my emails. Great. But what if I want you to now look through all my emails, look through all my Slack messages, and then I want you to go in and also look at a journal paper that I'm writing. And I want you to look at the emails where I got some response

14:31 from somebody. And I want you to take that, go into, go into that. Oh, and I also want you to go and create some graphs for me. So can you also create some Python graphs using, you know, map pot, Lib, etcetera. And I want you to do that all for me autonomously. What I'm describing to you now is absolutely achievable right this second using cloud code or codecs. It is also possible for the tools that that that that NLM is essentially calling or the agent is calling can be simulation tools. So if I extend that further and I say I want you to be helping me to do some journal paper, maybe I could say now, well, could you actually, could you

15:16 run some simulations for me? I have an idea. You know, here's some e-mail correspondence, here is a thread on a Slack channel with some collaborators. Not just write the paper, but will you go and set up an open phone simulation that that tries out some of those ideas we discussed and, and can you kick it off? And then when it's done, can you create the graphs and can you do it for me? That's completely achievable today. That is not future thinking. That is achievable today. My point is in what part of that then is this CFD tool? The main thing it's not is it because actually codecs or clawed or vibe code now with

16:02 mistrial entering as a new sort of player in this field from with a heavy sort of sovereign AI angle, European angle. So codecs clawed or or vibe, they can actually go and do that for you. And you don't have to be an expert user of Open phone because most likely those models have either been trained on sufficient smart data that included Open Phone documentation and various reports across the web. Or more likely that in a real life scenario, and maybe this is where the commercial bit comes in, one of those providers could have specifically added that capability. They could have looked at science and engineering as a key area they want to get into, and

16:46 so they will actually try to make their models good at doing this. At that point, the stickiness or the the difficulty of using ACFD code is no longer there, which is arguably one of the reasons those codes exist or have such a strong Moat because they're not easy to use. They require expertise, they're very well validated. If agent starts to come in and very powerful LLMS come in that can use these tools very well, then actually the uplift of changing to a different tool is not as much. And the ability to orchestrate multiple tools is also the same. So it raises a very interesting question, which is are the

17:34 software companies just going to reside to being just tools and skills and actually your relationship will be more with the open AIS, the mistrials, the anthropics, etcetera. Or, and it is the case today, will new companies arise that are trying to do this? Interestingly, and there's very little public information on this, but you have companies like Prometheus, I think Jeff Bezos put out some announcement about wanting to create a artificial engineer out of this. So read into that what you will. But it sounds pretty clear that they want to build some sort of capability like that. You see the recent acquisition of Meai by Mistral and you see a

18:17 huge interest from open AI and Anthropic in general, AI for science, which makes you look at the future and start to think there could be a new wave now of companies coming around. And there are companies that are now base themselves purely to create agents to create artificial AI engineers. This in some ways has the opportunity to massively democratize and and reduce the need for you to be some specialist. An example I gave is I remember when I was an engineer that being a CAD expert was a challenge. You would have to learn how to use NX or CTIA or SolidWorks, and frankly, if you didn't know how to use those, you couldn't use it.

19:04 You would have to go and find somebody, have to pay someone to do it. And I personally found it a blocker actually, because I knew what I wanted to do, I just couldn't do it. Now Fast forward and by the way, similar extent to some software tools where you know CFD, right? I understand the theory, I understand what I want to do. I just don't know how to use the code or I don't have to script it or whatever. As long as the Asian knows how to do it. I can say, you know, create this in this scan package, set up the simulation, this tool. It's kind of interesting then that it relies on a good skills capability. It relies on those tools being

19:45 open to agents. And that is a very interesting point, which is will companies embrace this, you know, agent first approach or will they try to put more guard rails and and even block people to use those tools via agents? Will there be some terms and conditions and licenses? And and that brings up another point around licenses. A lot of licenses are based around a human having a seat on a on a machine, and just like cloud with the ability to scale disrupted the per core license model. And really wanted it more to be like a just a power session like like Star has with his POD license. Finally, and I know this is rambling on a little bit, but

20:29 I'm just dumping all my thoughts down here, what about open code? Open source codes? Much of the challenge was around support and capability and validation. Is it not the case now with coding agents and the power of LLMS that these codes actually could become very capable and and could be used more easily in a workflow instead of a commercial tool because the agent can write and script and validate and go off and do test suites as much as possible? So I'm not saying at all that you will see a complete wipe out of all the current CAE companies. In fact, it definitely won't be like that, but I just put this out that there is a very interesting opening it because

21:19 these two new technologies around sort of foundation models. This idea that you could have simulations that would run in seconds or minutes instead of many hours. The idea that if you have such surrogate models and foundation models tied with agents, you could finally realize the sort of reverse engineering or design optimization, the automatic generation of cases. If you have a pipeline that is able to do that, then you could simply have agents go and creating optimum geometries for you, running those simulations autonomously with you being essentially the the orchestrator. The potential for productivity is absolutely massive.

22:02 It is. It could finally realize some of the dreams of people in the CAE and the EDA sector around productivity boosts, around finding designs around people with just domain knowledge driving the code rather than code knowledge. And that is important from an education point of view. I would say what matters more now is not are you really good at writing Python? No. Are you really good at using Openfox? No. Do you understand physics? Do you understand numerical? But do you understand how to pose the prompt? Do that is more important? Do you almost then it starts to become that you need to go higher up. If it's so easy to run ACFD code, why don't why don't I

22:45 start to run an FEA code as well? Many companies are you are split if you are ACFD person or an FEA person. You're split if you're an acoustics person. Well, now you could become more than engineer who simply is just driving the tools. That position in a company is often the sort of chief engineer. Well, maybe now it's more important that you understand trade-offs and and larger engineering concepts and doing this sort of specialist deep down analysis can be more automated and clearly simplifying what it. There are many industries where this will happen for many years, but I'm I'm sort of trying to Fast forward and play a what if scenario to, to to all this

23:27 point. It raises many interesting questions. Therefore, on the if the LLMS and the agents are key. It raises questions about traceability around sovereignty. Who owns these models? Where these models done? Should you be developing your own models? Can you take them off the off the shelf? So many, so many questions. It, it rises, but I, I find it a fascinating area that I think will be a key technology driver, an innovation area for, for, for years to come and something that I'd like to explore over the course of this, of this season. I want to look at speak to some people who can give insights into this on the the the core

24:12 ideas around sort of foundation models. Because the reason that foundation models of surrogates are so important is that if a tool still takes a day to run the the the agentic sort of optimization workflows will ultimately be massively Hanford by the actual tool itself. If the tool can match the speed of the agent, the LM, then that is what will really have the breakthrough. So whilst kind of the advantage of the agent is that, and so as you don't need to use any of these surrogate model technologies, they're even more optimum when you when you do. And, and so this is where the the, the combination comes from. So, yeah, I don't know what the future will, will, will, will

25:01 be, but I know that it is certainly an exciting one. I think this is a great time to be in the start up space. It's a great time to be researching, to be in academia, thinking of your PhD, thinking about what topics you want to do. It's a challenging time to be an end engineering customer. You don't know what to pick. It's a challenging time to be an established ISP, but it's also a great opportunity because you can potentially help drive your company to bring out the next, you know, industry 4 point O or five point O or whatever the the number is today. So I will, I'll leave it there. I'm sure I didn't cover many

25:41 things and I hope you appreciate that this isn't me in any way diminishing what will be the classic example of any new technology taking years and years and years to them. So I'm not saying next year or even 5 years everyone will be doing this, just like some people are still running on Windows 3.1 or Windows 95 in some places around the world. But I am confident, I am definitely confident that if we look in three years time that the role of agents and, and and surrogate models on AI will be transformative. And so please, if you're listening, if you are one of these people who are still sceptic of AII would respectfully encourage you to to try it.

26:28 And I would hope that you will believe me. The final call out, and I will say though, is in order to convince the community, we need transparency, we need openness as much as possible within a commercial mindset. People who know me know that I have always tried throughout my career to, to do things in a transparent and open way. It's part of the reason for doing this podcast is to sort of share the knowledge. The reason that I've created these sort of open source data sets is that desire to do it. The reason for doing these workshops or do CFD and others is to try and share it. I am fully appreciative of the need for commercialization. It's how the world works.

27:08 But I think the one key thing if this agents and AI and everything is going to be a transformative. People will only get on board if they can believe it, if they can see it, if they can see transparency, they can see publications. I feel like academia and scientific rigor has a massive part to play in this transformation. And if it can be done with scientific rigor at transparency and good science, then I think the the world will move on. If it's done behind closed doors with no publications, with just sort of a, you know, black box solution that that will not convince. And in fact, that will in some ways be a barrier to this, what

27:46 I feel is transformative technology helping. So with that, I will leave it there. I look forward to hearing your comments and I hope you'll join me on this new season. Please listen to the old episodes. There's quite a few, quite so many hours, but I hope you'll join for this season. So with that, thanks for listening.