Five tips for CAE engineers in the era of AI

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 in Formula One to some of the world's top academics to understand how fluid dynamics, machine learning, supercomputing are bringing in a new era 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. So today I wanted to give 5 tips

0:45 for CAE engineers in the era of AI. Five things that I think will be useful for you from a career point of view and from hopefully making the most of what I think is quite an exciting new technology. You may disagree with some of these things, and if you do or have other tips that you think people should should adopt, then you know, feel free to leave a comment if it's on YouTube or LinkedIn or wherever you're listening to this. OK, so let's get started. First of all, I'd say an an open mind would be tip one. It's very easy to come from a negative judgmental viewpoint when it comes to AI and it's natural. We as humans often, you know,

1:34 see things that are pushed at us and sometimes we have a temptation to, you know, be skeptical of some of this. But I think that is probably the the worst attitude you can have. I think having an open mind and being willing to look and listen and read is important. At the end you may still make the same judgement as you did at the beginning. But I think most people who actually look into this do end up changing their mind and and they understand better where some of maybe the marketing or hype is over egged and but where there are actual benefits. It is very tempting, of course, if you have 30 years experience, you know, like some people have

2:21 in in CFD or FEA, to almost be insulted by some young, you know, 20 year olds doing ML research and suggesting that their method, you know, can be better. But sometimes that's from a good intent place, they're excited and maybe they don't know or they're not from the CE domain. And, and so having an open mind, and this is true on sort of both sides on the CEA side and the ML community is important. The second one, which you sort of need the first one to get to the second one is educating yourself. You know, a lot of companies call this continual professional development. And I think this is very important in the era of AI. The problem and I which I think

3:14 creates sometimes this push back is because the, the ML community and the CAE community come from very different backgrounds. They often have studied different courses at university. They, they have a different preference in terms of software, in terms of programming. They, they're, they're two different communities and the challenge for CAE to look at AI is that it all seems quite foreign and complex. And This is why I think educating yourself is important. Going straight into a journal paper describing an ML architecture may be a bit overwhelming. It certainly was for me at the beginning, but there's a lot of content now that helps you. So Coursera is one of those

3:58 platforms where they have some great courses by lots and lots of people from different, you know, walks of life and different backgrounds that can really help you. Your company may already have a subscription or if you don't, you know, like I did it in the past, I think paying it yourself is a worthwhile investment. But beyond these sort of certifications like ABS does them, other tech companies do them. A lot of conferences, especially with AI topics, have a real push around transparency. And so I found that many of those conferences actually published their entire talks online, which is fantastic because you can actually watch on YouTube or whatever platform

4:39 you prefer Many of these talks, you can speed them up. You know, if you'd want to get through quickly, you can also use AI. Remember to summarize papers. This is something I do a lot. So for example, if you have a, you know, a paper like attention is all you need or on Transformers or maybe something specifically on CAE related AI. Now, most of these AI large language models can do a pretty good job of explaining things. And if you say, if you upload the paper and say, please summarize the this to a lay audience or explain to me why how this bit works. Or can you can you give me more examples of this? It's amazing how these AI can actually help teach you to learn

5:28 AI, but you can't get away from sometimes speaking to people. And that's one reason I would definitely try and broaden yourself and go maybe to even AI Pacific conferences like Nurips and others. Come and go to some of the workshops, listen in. You might not understand everything, but you'll slowly get and build the network. You'll start the people one-on-one and ask stupid questions. People who know me, I often I'm in meetings and I'll say, I'm sorry to ask a stupid question, but and often times it maybe not. It's a completely stupid question. And regardless, they usually explain something in a easy to understand manner that it would be difficult for me to get.

6:09 And sometimes all you needed is a one or two key concepts. And once you get the concept, you're like, OK, I see it now where at the beginning it it, yeah, seems a bit foreign. And the final ways podcasts, obviously I'm biased, don't make my own, but there's a lot of people creating them now, interviewing people and those again could be a a great thing. Finally, of course, depending on, you know, where you are in your career, going back to university or taking the university course could also be a worthwhile investment, particularly if you're thinking of a, a job change. You know, if you actually want to go into AI and say the sort

6:44 of two roles, you can either be the sort of CAE domain specific person, or you might want to be more hardcore into the AI itself. And probably for that you would value from some of these courses. Although many universities now do offer these online courses as well, like I think Stanford does them. So you know, there's a lot of options now. And I don't think there's any excuse not to learn. You know, it's not like 40 years ago where you have to go to a library where you're in a physical thing. The Internet gives you so many opportunities now. So if you have an open mind, which is 1 and you've, you know, educated yourself, which is number 2, and of course they're

7:22 continual processes, then let's talk about the AI physics. And this is probably the most common use case or one of the most common use cases for AI in the context of of CAE and engineering AI physics, you know, referring to the use of AI to create typically some sort of surrogate model where you take data to train a model. And then once the model's trained that inference, you can give it a new condition, a geometry, a binary condition, and they'll go and predict it, typically in close to real time. There are other use cases people sometimes look at developing better models, you know, transition model, the turbulence model using AI.

8:05 But I'd say probably the predominant one that you may come across is more of the surrogate modelling. And yes, some people say, oh, we've been doing that for decades, we've produced all the models, but typically those did not have the flexibility like modern day ones in terms of the non parametric ability as you can just bring any arbitrary geometry or flow condition in, if you've trained across it, you can get the results out. So things are different today than what they used to be. And that common argument of, well, we've just been doing ML for 30 years and they've just changed it to something else is probably going back to #1 and #2

8:43 and open minded educating yourself to realize, no, no, things have changed. Although of course they're based on foundations from from earlier. So how do you prepare yourself for the air physics? Before we talk about it in more detail, I said the first one is, is data Today, the biggest difference between your probably common use of AI with Google Gemini or ChatGPT or whatever is they're already trained. You're essentially just doing a text prompt. You're doing inference on a pre trained model. The big thing with AI in the context of CAE and CFD and FEA is you're probably going to have to train the model yourself. That's most likely at least in

9:22 the short term. So you need data and if you have existing data, it's what formats it in. Where is it at? Can you get it to the model? Have you scrutinized it? How? How do you describe it? The model needs to know what it was generated with. Not all data is identical, and that is often the biggest challenge. Did you run two different geometries, but one, you change the CFD settings or did you change the material properties for that crash test? So labeling the data, classifying the data in a format that a model could read is important. So for example, you know, coming up with some sort of schema, some Jason format where you can

10:03 say, OK, this data was created on this day using these settings by this person. You may even add security into it. I want this to be able to be trained by this model, not to be trained, you know, if you have different sorts of data that you want the model to know about, you should try and describe as well as possible. The other one, and this is probably more translating now to new data, is the data that you would traditionally keep or destroy should be reconsidered in the era of AI. The example I would give is some CFD. You may say all I need out of it is the drag and the lift and some pictures of the flow. And after a certain point, why

10:44 do I need the full 3D volume that's 50 gigabytes. I'm just going to delete that. But by deleting it, you've probably lost a lot of information that would be needed by the AI model to go and train the 3D volume solution. It needs it to be able to learn 3D field. And if you deleted that, well, you are going to have to regenerate it. But then how do you regenerate it with that version of the software that was done and the, of course, the cost. So traditionally this was a balance of storage cost versus need. I would argue now, because data is the key part of AI, it's the biggest cost of AI when it comes to CAE, you need to keep the

11:26 data. So I would argue it's best investing in paying more for storage to have the data. Even if you're not today trading AI models, you will do, I'm sure at some point. So generating data, thinking about how you keep it and maybe outputting more information than you think you need because the AI model may need it. So if normally you would just save the pressure and the velocity, you might think, well, what about all the other variables that I might need beyond just what I'm traditionally getting out? You know, instead of being just the stream wise velocity, maybe I need all the components of the velocity, little things like that that you would say, well, I

12:07 don't need it because I don't need to explore it. You might need to train a model later. Another angle to look at is monetizing this. So if you're a business or engineering company or even individual, you may find that your data is very valuable and you may consider is the way of making a value at that. Can you change your business around a little bit to offer this data, whether it's computational or experimental, to help people train? If you're a company that operates a wind tunnel, maybe you start to think about using that to generate data or to monetize your data. So there's a lot of opportunities in the a, in the

12:46 era of AI physics, which is true for AI, for science, that data is key. And if you have data, it could be valuable. And you really need to have a good strategy of how to deal with this. OK, So you have an open mind, you've educated yourself, you're preparing for the AI physics. Now what about doing it? And I think the argument or the thing I would like to discuss now is really about the builders buy it's a common question I get when I speak to many people is around, you know, is or an off the shelf solution. Should I be building this myself? Well, obviously linked to #2 you need to educate yourself. And I should say sorry on #3

13:27 actually, one thing I forgot to say was staff. You can prepare yourself from a technology point of view in terms of getting the data ready, processes ready, but you need to prepare yourself in terms of hiring. Do you have somebody in your business who is it an expert? If you're a manager, if you're an individual, the the preparing is to prepare yourself, you know, by doing the education. But if you're a company, you know, don't expect that people who have been doing CFD for for 20 years are going to be the best people to do AI and help you develop. AI is probably going to be, you know, a bunch of graduates who have AI native and are very familiar with those programming

14:06 styles with that, maybe they've studied at university. There's just a reality that staffing is important. And you may also need to look beyond your typical recruiting grounds to get those staff. Those people may not come out of the universities that you would normally go to, to require to hire engineers. And I think this is really interesting blend. And it's a career opportunity that AI is needed for engineering, but also engineers are needing it are needed at AI companies. So OK, let's go to #4 So you've don't open mind, educate yourself and you're sort of starting to prepare yourself from a data staffing. What about the actual doing it

14:47 now? Well, this is the bill versus buy. And I would put the analogy to something I'd probably know best, which is CFD. And if you looked in the 70s and the 80s, there weren't really commercial solutions on the market that were the de facto solutions. In fact, before the 1970s, people, there was no commercial solutions. And so you had to develop the code to yourself. And that was a big thing of NASA, but not just NASA, but almost all aerospace companies at car companies to some point manufacturing, they would develop their own codes out of necessity if they needed to. There wasn't, you know, there wasn't a commercial solution available.

15:28 It was only in the 70s and the 80s and certainly going into the 90s and the 2000s where commercial codes became far more mature. Interestingly, if you look today and this is maybe contentious opinion, lots of these companies that did traditionally only use internal codes are starting to use commercial codes and that shift is accelerating because frankly, the commercial codes are getting so good. They've hired so many people, they've, you know, acquired start-ups that is your code really as good as theirs. I think it's a interesting debate. So debate of do you see it as core to your mission to have your own code or actually is your main business to go and

16:17 build something and the code itself should be done by somebody else. Most people are starting to shift towards the latter, that actually it's better just to buy in the code and have less people developing their own code. Now again, I'm not saying that is the right approach, I'm just saying that is what I see happening. So why am I saying that? If you look at AI physics, it's probably the fact that we are in the equivalent of the 70s or 80s. We're in this new phase where there is not the same maturity of commercial solutions. It's largely start-ups at the moment. And you could argue that actually, could I build this

16:57 myself? Does that give me a strategic advantage like it was a strategic advantage building your own CFD code? I would say I can understand the arguments for both. I can today I can see why maybe you think if you hire some, you know, smart engineers and build an open source frameworks that you, you could do that, but you have to be prepared to keep up. And that's my advice or warning that even if today there is a split on, oh, it's not so obvious whether to build or buy, you should be flexible and prepared that maybe in two years, maybe in five years, maybe at the extreme 10 years, it's I think it highly like that

17:42 the commercial solutions will be as good or better than what you could do yourself. And so just as people are now having to debate whether to bring in commercial solutions from a non AI point of view, I would say you need to be flexible enough to be able to adopt when there is a good off the shelf solution that's maybe better than what you could do internally. And that's part of the sort of prepare yourself and planning. Luckily today most solutions are sort of API driven and therefore if done in the right way, you can sort of integrate things together a bit like you can use the API of open AI to call that model and call the different

18:20 model. I suspect it'll be the case. So I would 100% advocate people coding and building stuff themselves, whether they go to production with it or they use a start up or or buy something. It's very much an individual choice. But I do have a strong feeling that given the investment it is likely that they'll be so many better and good commercial solutions. Given how big this AL market is for for CAE, that you may need to be flexible on that choice and reassess it and don't get locked in to thinking I can't take a commercial because I've, you know, gone down building myself. I need to. There's nothing wrong with having a mixture. And finally, .5, which is

19:09 probably the newest, I would argue in the wave of AI for CAE is agentic AI and a bit like AI physics. Sometimes people can get people's backs up with the marketing and the way that it's bullshit this that AI can automate and do everything. Essentially the agentic AI is to sort of next wave of where LLMS were. So in the sense what are the frustrations of using a traditional LLM is it can't go and do things for you. It's very much a, you know, look at this document and summarize it for me. But what you would really like to do, and to be fair, even now, some of the off the shelf that can do it, you'd say, I would like you to go and do this for

19:52 me. Go and research something, go go on the Internet, go and search for this, then call this, then do this. So what does that mean in the context of, of of CAE? That means at it's very simplest things like a text prompt to set up simulations and run simulations. So rather than you clicking buttons and and going and setting up a simulation and manually running a bash script. It's through a text prompt. You should be able to instruct an agent. And usually there's some sort of master agent that's then controlling other agents and sending that on. And I will do a dedicated episode on this because I think it's such a fascinating topic that essentially allows you then

20:34 to quote, UN quote, have a sort of AI engineer where that agent would be able to go and call other agents. And again, the reason I'll do a dedicated episode on this is there's been some good papers out there that I'd like to discuss and and talk about. But at a very high level, it essentially would be the ability through a text prompt for one agent to call another agent, which is perhaps a surrogate model to go run a simulation. But in importantly, another agent would then perhaps do an analysis of that, but in a fully automated way. And then finally, another agent may decide to go and then do some optimization of the

21:15 geometry and pass the information back to your surrogate model agent. Then do more analysis, let's say image analysis through an LLM and they may write you APDF. And the important bit that's all been kicked off by one prompt from yourself. So it's agents that are essentially calling each other. And there's a lot of complexity, of course, with this, but it really gets much more than just the surrogate. The surrogate still relies you as a human to run the simulation, a bit like ACFD simulation or FEA simulation. The agentic side really starts to get more, I would say, into the the vision of AI, where it's more fully automated. And if you extract this to its

21:59 maximum, you can imagine many Asians acting almost like their own engineering company. That's obviously quite far into the future. But I would start preparing for that. I would start reading up on that. This is very much where the the AI researchers at the moment go and read papers, type in into Google agentic AI, go into YouTube, watch videos, look for start-ups, speak to them. This is definitely a new wave that's coming that AI physics is a crucial part of it. But the agentic AI is arguably more potentially transformative and more important to to be aware of. But there's even, I'd say less solutions on the market now,

22:38 which is why it's really good to stay ahead of the curve. So those are just 5 tips being open minded, educate yourself, prepare for this sort of AI physics revolution. Get involved in the AI visit tries to foul build by, you know, kick the tires. And then prepare yourself for the agentic AI move, which could really blow away and be quite transformative if things live up to what people hope for. So I hope this has been interesting. I tried to keep it a bit short and snappy. I hope you've learned something. Agree with at least some of the stuff I said. And what I'm going to do is over the course of the the next season, we'll, we'll dive into a

23:21 couple of these topics a little bit more and hopefully do #2 educate a little bit. So with that, thanks very much for listening and hope to see you in the next episode.