0:15 Okay, hi guys. My name’s Jon Wilde and I’m going to talk to you about advanced workflows inside simulation. I also think that I was given I was given the longest title ever. So really, I’m going to talk about AI in engineering. And I watched a few presentations over the last couple last yeah, day and a half and I figured nobody actually talks about who they are first, so you don’t know who you’re listening to.

0:43 Going to do it a bit different even two slides just to see how this goes. Just bear with me. Maybe it’s good, maybe not. Hello there. How do you do that? How was you? Okay, beautiful. No, that was a spoiler. Okay, so I run the product team at SimScale and I’m a mechanical engineer by trade. My first car was what we would call a shitbox in the UK and I had to learn engineering the hard way through this because it broke all the time.

1:12 I had no money and I had to fix it constantly. So I learned engineering at university but this really forced me to get my hands dirty. Then I built my own one which was like a rocket and this was awesome. And then I had another one in Germany where I think I’m only sharing it because it cost me 12 euros to have what has to be the coolest number plate for a 205.

1:34 I’ve been at SimScale for 8 years running the product team. I’m based in Portsmouth in the UK. If you don’t know the band Portishead, they named themselves after our town because it was the dirtiest like horrible most most horrible town ever and they’re really depressing trip hop band. It’s nice now but check out the band if you don’t know it. It’s just here on the estuary so we can see whales across the way.

1:59 I live at the top, and I’m often rowing out on the sea. Feel free to come talk to me about SimScale, about AI, product management, even controversial pizza toppings, cuz I know it’s a thing between UK and everyone else about pineapple. Just come find me and we can talk about it. And I love what we’re doing at SimScale because we’re trying to do something hard. There’s definitely easier ways to make money, to be honest.

2:22 But we’re trying to change engineering and give everybody to simulation. And go against the big guys essentially, you know. Okay, now I’ll get started. Thanks for humoring me. I appreciate it. So, I’m going to talk about four things. To start with what I’ve heard from engineers over the last couple of years. Then the current state of AI and engineering as we’ve seen it so far. We interviewed hundreds of different engineering leaders.

2:48 Then I’ll talk about the problem we’re trying to solve. And then how we’re solving it, and I’ll talk about physics AI. So, the instantaneous prediction of physics. And I’ll talk about engineering AI, which is using agentic workflows to do more holistic high-level workflows like up design optimizations. But, I’ll get more into detail with those in a second. So, this is what we’re hearing from engineers that they don’t want like you just said, they don’t want probabilistic answers, they want certainty.

3:16 They need to show value, and they need to be confident in the AI they’re using. They need to either make their designs better or get to a good design faster. And we hear it all the time, too. They don’t want to be replaced. I don’t think myself this is a risk because we need these experts in the loop. Like, whatever the AI’s doing, it might be awesome, but we still need the expert to validate it and put a human stamp on it at the end, right?

3:41 And engineering leaders have said different things. So, they said they need to accelerate development at all costs, and can AI help? They can’t afford to grow their teams. Can they use AI to do it? Another leader, he’s actually here, so he might recognize this, told me a couple of weeks ago that he, by the end of this year, has to not only have humans reporting to him, but agentic agents as well.

4:01 So, his team needs to be comprised of both, which I thought was pretty interesting. And it’s great because a leader is encouraging this, right? And another, who made me sad, to be honest, said, AI didn’t work. And now it’s going to be really hard for him to convince his manager to experiment again. So, I think it’s our goal to educate engineers to figure out how to experiment with AI, like at least starting small, and then later on finding the huge kind of value that can be unlocked.

4:32 The current status, I thought I’d say, so we interviewed 350 engineering leaders, and there’s only four slides. I’ll call it fun with growth, but we all know they’re not that fun. Just to give you an insight into what we’ve heard, cuz I think it’s interesting. So, only 9% have so far a mature AI, rollout. You could also say 9% is actually quite a lot, right? If 9% of people are using AI confidently, that’s pretty cool.

80% are, experimenting with pilots, and this is also good, but my only concern is, are they going to get stuck there? So, are they doing the right things, or are they going to find they’re going to get burned as well, and then not be confident to try it, cuz we need it for this kind of large unlock? They’re able to work or they’re able to explore three times the design space they were before by leveraging AI.

5:22 So, this is a thing really cool, but I think a lot of this is to do with physics AI, and not engineering AI. And the leaders are going way further, so they’re exploring massive design spaces already. Some of our customers are doing this, too, but they’re able to turn around RFQs, so requests for quotes, for requests for proposals, in about three times faster than they were before.

5:44 I still think there’s a long way to go. I mean, if you look at the numbers, they’re kind of moving from weeks to days, but I think we can get down to minutes. I’m really sure it’s like you just said, right? It’s It’s really the way to do it. Modern infrastructure is critical, but I thought actually more critical is buying from leadership. So, everyone I’ve seen who’s moving and leveraging AI better is doing it because their manager is pulling them into it.

6:09 They can see that there’s a benefit there and they’ve got that support to learn and to experiment. And while data seen as the biggest blocker, I still think that’s because we’re thinking about surrogates, we’re thinking about physics AI, but if we think about agentic workflow and engineering AI, we don’t need all that data. So, although this seems to be the largest blocker, I think it’s cuz we’re all thinking about it wrong and there is a really large unlock.

6:34 We just need to find out how to do it. So, the problem we want to solve, like is I would say already solved in software. So, I think probably everyone in here, we’ve seen examples already, is by coding. Like, I can’t code, but now I suddenly can, right? I can make my own everything. And it’s amazing. I’m so much more efficient just through board, to be honest.

6:58 So, I think in software, it’s working beautifully. But it’s not yet working in hardware. Where and our customers are saying the same, the ones that aren’t using AI at least, they’re still taking weeks to do one iteration, sometimes longer. And this is where I think AI will have a really large impact when it starts to work inside a physical world and not just inside software. Part of the problem, I think, is how historically engineering has been structured.

7:30 So, inside companies, you’ve got simulation experts they’re working with some CAE tool. They probably got a workstation under their desk and whatever they do probably stays there forever. They’re doing some kind of pre-processing, some post-processing, already like three different tools. Maybe it’s going to a cluster, the data stays there. Maybe they’re using some scripts to automate, but again, where’s the data going to reside? And then upstream, you’ve got the design engineers, they’re producing the CAD models.

7:56 Also, hopefully integrates into PLM, but once it goes to the simulation engineers, it’s already disconnected and it doesn’t get passed back quite often. And we’re wasting a lot of time because of this siloed kind of manual process. And we all know that like programs can take years to get from idea to something produced because of all the wasted time that we that we have here. And it’s really difficult for AI to enter something so complex.

8:20 And what we want is this, right? To say, “Hey, I want a pump. Here’s my requirements. This is the pressure head that it must give me. This is the envelope it needs to operate within.” Beautiful. But it doesn’t work. Because the problem is difficult. It’s really kind of multi-dimensional. And the data that that we need to access is often locked inside organizations and it’s spread about all over the place.

8:44 So, this is the problem that we need to try to solve. And at the moment, I’ve seen no actual results from it from anybody, to be honest. If we take a big step back and think about the design cycle, this is really basic, right? But just thinking at the beginning, you’ve got a designer, they’ve designed a manifold. After some simulation lead time, which normally is some kind of waiting time, there’s a simulation cycle time where the expert, the simulation person, will do the simulation.

9:11 So, they’ll find out how this how this manifold performs and does it meet the requirements. But they’re normally working quite separately. I think there are gains to be made in the simulation cycle time. We’ve heard about it a lot already, especially around physics AI. But I mean, just GPU acceleration or maybe quantum computing, too, are different ways to speed this up. Or shifting more work upstream. Even though simulation is kind of democratized, so they’re already doing the really hard stuff.

9:36 It’s hard to give more stuff to the designers. I think the only realistic solution now is physics AI. So, these instantaneous predictions. But, I think we’re wasting almost all of the time in the simulation lead time. Where people are sending emails from one to another. We’re waiting for the engineers to have time to do the simulations. And this is where the gains are the greatest. And this is where we can use engineering AI.

10:08 These are some examples from people that we’ve spoken to. And just for one iteration to go from one CAD model to one simulation, it’s taking a couple of months. Somebody yesterday said it took 18 months to go from an idea that they had in CAD to a simulation, which I don’t know. It’s crazy to me. It should take seconds, right? You need that insight as soon as you have the idea and you’ve got the design, you should be able to get the insight and understand how it performs.

10:35 So, this is where engineering AI comes into play. Because it can automate that process. It can break down the silos that we have at the beginning, and it can compress the simulation lead time. I think that’s where the gold is, and that’s where we need to start working. So, teaching AI to engineer is actually a bit more complicated because the process looks a little bit different. I don’t think we’re ever going to get to the point where we can zero-shot a design from an idea to the final thing.

11:03 We’re always going to still go through these iterations. You’re already starting from something that’s probably in production, and you need to iterate more. So, you’re going from a design to a simulation. Doesn’t work, back to the design, make some more changes, back to another simulation. Maybe this time it works. At some point you get to the point where it’s good enough and it meets the requirement. And here we need engineering AI to do this for us.

11:25 So, it can automate this iterative process, which I think we’re never going to escape from. Which is fine if we can use agentic workflows to do it for us. Engineering AI can leverage physics AI to do the instantaneous predictions to accelerate this process. And if you’re using SimScale, you’ve got access to totally broad physics. All of your data is together, which I think is beautiful because the designers can do simulations using agentic workflows.

11:51 And because it’s in the cloud, the experts can just use the URL, jump into the project, so they can collaborate in one place. There’s no kind of, data spread about across lots of different, sites. So, I think by combining engineering AI and physics AI, we can start iterate and we can get to, a final solution really nicely. So, just focusing on physics AI for a second, then engineering AI afterwards.

12:23 So, all of these things are possible with physics AI. If you’ve got enough data and if you can train a model, the designers can easily do all this, so you don’t need to be a simulation expert. And as long as a simulation expert created the model in the first place and has good confidence, upstream, someone can do all of these simulations. So, the problem is you need the data to be ready.

12:46 If it is, though, you can go further and you can do, really nice, easy, quick design explorations like this. So, it doesn’t take any time at all if you’re using physics AI to step through lots of different iteration. And you can even work backwards. So, here we are using a foundation model. This is trained with physics Nemo. Then we’re starting with the requirements and really quickly working backwards.

13:08 You can’t see cuz the one was listening to some funky music last night. That’s the zoom. They said AI. I think they just listened to the two of us. Well, there you go. It took 5 minutes, anyway, in real time to go from requirements to find the pump design by using the foundation model. Okay, engineering AI works a bit differently. So, here we have an agent that can work inside our SimScale.

13:37 It understands what it’s looking at and here I asked it to set this model up with the loads on the top surface. It did that for me, so it understood where the top was. It constrained the bottom and it run the simulation for me. It took a little bit longer than that. I obviously cut a couple of moments out. But, it was great. I posted this on LinkedIn and somebody said to me they kind of called BS.

14:06 They said, “Come on, it doesn’t really know what it’s doing, right?” I thought maybe it does. So, I asked it like could this structure be used on an aircraft because that’s what the guy asked me. And the agent said, no, that’s stupid. It’s too heavy. It looks like ground equipment. You haven’t taken it for vibration and here are the resonant frequencies that you need to consider if it’s going to be going on an aircraft.

14:27 And but honestly, it looked like a test stand. True. So, this is awesome. So, if if I didn’t know what I was doing, I would be informed by using the agentic workflow that we have inside SimScale. So, to put it together, we have a deeply knowledgeable agent that understands what it can see, it understands the simulation. You can obviously train multiple simulation specialists if you want to so they can understand specific applications.

14:58 And you can layer physics AI on top if you want to for the rapid kind of optimization. It can also work outside of SimScale, so it can work inside traditional tools, which is where people are generally most comfortable because that’s where they’ve spent the last few decades working. So, they can layer that our agentic workflows on top of what they already have in place. So, I’ll show you how it works.

15:25 On the left, we’ve got a server rack with a few different options. You can do the math really quickly, but there’s a lot of different permutations just from a few different options of heatsink designs and fin placements. And I had the AI run through it with me. So, I started off just basic. I gave it a single CAD model and it understood what it knows this kind of model, so it’s seen them before.

15:53 And I asked it to run a thermal performance for me. All the fans are names from the brand that they come from, so it knew what was what. It could set them up with the correct flow rate. This is set up a funny way. I’m sorry, it’s a little bit flashy. But it set up the right heat loads, the correct materials. So, by the time it was ready to be simulated, it was ready to be simulated like it was existing in the real world.

16:14 So, it would hopefully perform as it would in the real world. Once it was done, you can ask it not to do this actually, but I wanted to. It will ask you for a one final validation. It’s going to use some core hours, right? And sometimes you want to just pause and check, but you don’t have to. It’ll go away, it’ll create a mesh, and it’ll run a simulation.

16:38 And then once it’s done, we can check out some results. And what we wanted to do was understand what the temperature was on the chip, which was our main component that sits at the front. So, here you can see the airflow going through. The white are the hot components. The blue is the cooler components. And all I did is just ask what the temperature was on the chip.

17:00 So, all this is done through a Gentex workflows, right? Which is really nice. So, anyone could have done this, but the specialist would have told the AI what to do in advance, so it knew how to set things up. Leveraging physics AI, similar. So, the workflow on the left is the same. Everything is the same, but we can get instantaneous predictions instead. So, you get the same result.

17:23 And if you want to, what I do in a second is change the CAD to a different design. And again, get another instantaneous prediction. But then, what we can do is combine the two things. So, engineering AI can do the automation, and it can leverage physics AI in the background for the instantaneous predictions. Which means it can go through all the permutations really quickly to find me the best design.

17:58 Okay, just to wrap up, I’ll share a couple of things our customers are doing cuz I think it’s interesting to see that it’s not there’s vaporware this thing this and customers are leveraging it now. Combion had an interesting problem. I can’t actually show the CAD, but at least we can see that they were trying to recover these gases from a horrible environment. Imagine a heat exchanger. It looked very much like this.

18:21 And it was taking them months to kind of explore designs and find a valid solution. We helped them train a physics AI model. It looked a bit like this, but it was not this. And they could find the best options. So, instead of taking months, now they get to the design that they need in an hour. And what was really cool was that they found non-intuitive geometry.

18:41 So, they would never have not never. Maybe not never. They wouldn’t have really found this crazy design without using AI. Which was actually half the volume that their old model was. So, they could manufacture it for less, and it performed better. And it took them 1 hour to find. And now, they’ve got this model deployed throughout all of the designers. So, every time they make a new design, they can really quickly get an insight to see how does it perform.

19:13 And there are other workflows we can do too. So, you could upload a CAD model like this into Synapse and say like run different types of analysis for me and the agent knows what to do and it will go and run them all for you. When it’s done it can give you some results or report. It can explore designs. It can optimize them. But we’ve seen that already.

19:35 And on the right, I think this is interesting because we have a customer doing this today. They are they can take some kind of code or an RFQ. It can explore designs. It can explore the whole product portfolio. It can find the best solution and it can make a report and give it back ready to give you ready to give back to your customer. And this takes seconds.

19:54 And they were nowhere close to this before. They just didn’t know how to get from the vast array of models that they had to one that fit the requirements. And now it’s all fully automated. Okay? Thank you for listening. Appreciate your time. To learn more about the CFD fan computational design symposium, access the archive of previous presentations, interviews with speakers, and information about future events around the world, visit CDFAM.com.