Good. Hi everyone from my side. I’m David, co-founder and CEO of Simscale. Pleasure to be here. Duann, also the fact that he’s now leaving my presentation is not so cool, but that’s okay. And for the next 20 minutes, I’m going to talk a bit about our work in AI and how we’re seeing engineering teams you know, adopt AI in a practical way. It’s called a practical path to AI.

0:25 There’s many AI is a wide field. I’m sounds like I wasn’t the first one to talk about AI today. And the reason why we’re excited about AI is frankly we talk to a lot of engineers on a day-to-day basis and I think everybody here in the room to some extent cares about bits but also about atoms and we probably all became architects, engineers, scientists to sort of you know drive the the world forward to drive hardware innovation but more often than not as hardware engineers we kind of get stuck with this.

0:51 We join an engineering organization. By the way that’s not a political statement. I’m born in Germany, European, so you know this is not a political statement. We’re we’re sort of getting into the into the weeds of how modern engineering looks like into a larger engineering organization. Every product that you know got some level of complexity eventually has a big process around it and that’s what we end up stuck with.

18:51 And so, now we’re seeing all of these incredible AI systems emerge and we’re seeing, you know, everybody in the last couple of years had this aha moment of seeing a frontier model do really complicated knowledge work for you. I guess everybody, you know, is using them on a day-to-day basis and so why not have that in core engineering workflows? We’re talking to a lot of engineering leaders on a on a daily basis and all of them are trying to figure that out.

Sort of why is that not working? Why is core AI not working in in in engineering workflows and because it’s works so well in other spaces right we’ve heard earlier image generation coding you know all sorts of knowledge work has seen strong AI adoption why not core engineering workflows and in addition when we see AI in engineering it often times feels like this right you see and and by the way we’ve trained a lot of AI models so we’re also you know part of the problem here that the AI inference always looks great but you know What often times you don’t see is the AI training that happened behind.

No pictures there. I don’t want to sort of you know get problems with animal protection and whatnot. So so that’s really the problem. And why is that? Because frankly engineering as we all know is more complicated. I’m from Germany. Germans are obsessed with the V product development model. So no German presentation without the V product development model. But what it conveys is the complexity of putting a piece of hardware into the real world.

And so hence lots of modalities, lots of scale, lots of different disciplines, lots of things to get right in order to actually make AI work for you. And so that was for us the reason why we commissioned a small survey. We said we’re talking to so many engineering leaders on a weekly basis and all of them struggle with sort of figuring out how will it look like?

We just had nice conversations over lunch where other engineering software vendors see the same. So how will it happen? And I mean the the somewhat good news is that the charts are probably the small to read from the back I suppose. Anyways, I’ll tell you what what what it says. The the survey we commissioned, we asked essentially 300 engineering leaders about their expectations on how AI will enter the engineering workflow on what they’re seeing today, where they are today, etc.

19:08 And I think the two main takeaways were one thing that was encouraging from our point of view is we’ve been sort of talking with engineering leaders about AI entering simulation specifically since a number of years and I think the sentiment changed over the last two years. Everybody again had this aha moment with with frontier models and the sentiment changed. This is a survey like two two to three weeks old.

Many of them are actually now saying that says more than 90% are expecting a significant boost in productivity for their engineering workflows. Resulting from the adoption of AI yet 3% are seeing it and again from all the conversations I had here today I think the you know the statistics here everybody you know has the same statistics on this or sort of sees that anecdotal evidence on a week-to-eek basis surprise surprise as a cloudnative simulation software vendor the survey also said that cloudnative users are closer to AI adoption kept an obvious here but anyways you know that’s us and So what do we do with that?

We think the fact that you know this this complicated V model, the complicated the complexity we’re dealing with AI entering all of the you know disciplines of knowledge work but not yet core engineering workflows still leaves engineers often times sort of innovating in the dark. It’s a sentence we like because we’re a simulation company. We’re sort of helping engineers evaluate design decisions faster at scale in seconds.

And in most engineering processes that’s not the case yet. So an engineering team is left with trying to innovate in the dark. You’re making a design decision. There’s hundreds of engineers making design decisions. Yet there’s a very handful a very few like a very low number of engineers evaluating design decisions. And while somebody waits for the evaluation, they’re left innovating in the dark. That’s sort of a bit the you know we like the sentence and it describes a bit the the motivation we have at Simscale where we’re we think the technology is out there already.

Data technology is there to enable an engineering team to evaluate not a design decision in weeks but thousands and seconds. That’s a bold statement. And I hope sort of over a couple of stats about about Simskill here on the left and all what those stats shall show is that we’ve been fortunate enough to work with at this point thousands of engineering organizations to help them accelerate the way they evaluate design decisions by via multiple technologies.

And the way we think about it and here’s sort of you know so far I talked about AI very fuzzy sort of what is it and the way we think about it is that we think that there might be eventually you know large novel frontier models that do big chunks of of core engineering like in in one step right now that’s not yet visible but what is visible is what we believe is that every step of the engineering process sees the first AI tooling evolve.

By the way, this is not the V model as you can see. So, I hope you appreciate that I broke it down to a very simple model. So, we can call this the poor man’s V model maybe. And everybody that has done real design work also knows that it’s not that linear. Where an engineering team loses time is in the iterations. So you know you might it looks very linear and we’ve seen earlier I love Rubin’s talk and we we need to speak later where we’ve seen sort of AI tooling emerge on the left that accelerates the way design is being generated and then on the right side DFM checks simulations testing is sort of the role where you evaluate designs and today teams lose time when they try to iterate through those through those loops and that’s the big bottleneck and we’re excited we think it’s an exciting time in engineering because we’re seeing in each of those steps the first tooling evolve I would almost say in all of them and you know many companies are here we see Sinara down here shout out Andrew we’ll talk tomorrow is here there’s lots of companies doing great work around each of those steps and we think sort of if it starts to feel real that actually a lot of the mundane work in engineering workflows starts to at least be augmented at least accelerated eventually completely automated by AI systems and those working together and so the the place where we fit in because simcale is not everything to everyone.

The place where we fit in is that part in the middle where it’s about engineering and physics AI. Two AI systems that shall help to cut down the simulation lead time which we call it that is not just the actual solve time so that the solver is fast but all the manual work that it’s associated with actually setting up a simulation running the simulation evaluating the simulation those are the two systems that we’re building.

And just to double click on those and that’s sort of how you know how we’re going to market with this is if we if we think about that part of the engineering process. So the system architecture has been done requirements engineering has been done etc. And you come down to that you’re not actually going to mechanical design more detailed design work. You typically start with some intent with some parameters that you fixed.

There’s a manual process of somebody sitting down creating that design handing it over to the analysis team, methods team, whatever you want to call it. This is an iterative process. They iterate a couple of times through that process to eventually reach the final result. And the way we now think about that is everybody is familiar with running simulations takes time. So the the blue box on the right, the simulation part specifically when it comes to more complicated simulations, you’re going narrow into design space.

This is sometimes sort of the time waster, the time killer. This is where you know since years physics AI methods, deep learning circuit modeling techniques have been progressing. There’s you know many novel model architectures. They’re getting better by the day and we think they’re going to more and more, you know, be be spread wider, be used wider. But that’s not all, right? That’s sort of the the the pure solve part.

The other part is the fact again everybody that has run simulation knows how long it typically takes to actually set up that analysis. Doesn’t matter if you run it with a physics AI model or a normal solver. It takes a lot of manual time to set up the process to then run it to you know postprocess it to evaluate it yada yada yada. And that part from our point of view today the frontier systems can already do a lot of work around that given you have the right stack.

And then the way we organize ourselves is all of this is agnostic to the design methodology meaning we learned the hard way that it’s hard to change an authoring workflow of a company. Which is why we’re building these systems agnostic and so whatever the design approach is those systems flange to it. Good. Now let’s take a a closer look. Dan already said, I think I’m gonna be the first one that tries a live demo.

So let’s see how that goes. So far, thanks. Appreciate it. We’ll see. If it doesn’t work, it’s the internet here. That’s always my rule. So let’s take a closer look at Simscale. On the first, you know, on the first glance, yet another simulation software. So let’s take a closer look how customers are using it. It starts with a couple of a couple of different things. First cloud native meaning every you know all data that’s being generated all data that’s ingested is structured properly lives in the cloud accessible to the latest GPUs we cover with native simcale has about 10 solvers natively integrated five in CFD a handful in FA a bit of EMAC etc so there’s integrated solvers but it plays well with third party solvers as well covers a broad range of physics auto HPC provisioning so depending on you know the simulation you want to run or the training what to do.

HPC is provisioned automatically. And then lastly, again, I always find it’s funny that we’re an engineering software and it’s like a new thing that’s collaborative software kind of solved in every other software vertical but but in ours. So it’s out of the box collaborative meaning that engineering teams can deploy it out of the box broadly, right? So it’s not just a tool that sits somewhere in the lab, but it can be made available to the engineering team.

You can send out login. Then the next step is now that the stack is cloud-based you have basically a nicely structured semantics and syntax around the simulation setup around the simulation results you have nicely labeled data etc. You can start automating certain workflows. For that, we’ve woven the frontier models into sim scale and they are able to reason through simulation setups, they can reason through simulation results and the likes.

And then the last part physics AI the moment you then have enough simulation data gathered as a team and you want to go somewhere narrow into a design space, that’s when you can use a deep learning circuit modeling techniques where you train a physics AI model to go really really fast. Good. And with that, let’s take a look. I actually asked Duann if he can hold my microphone.

No, I was Okay, we can see Sim skill here live. All right. And can we see a little the little manifold? That should work, right? Okay. And just like a mini setup, Dre. Good. We didn’t practice that before, so I really appreciate it. So, what we’re going to do is we’ll simply have a an AI now take a look at this topology, take a look at this geometry, and actually try to set up a simulation.

So, let’s see if this works. Do we want to give an applause to Thanks. You can take a seat. Okay. Fair enough. So what we’re looking at here is like on the right this is this is really just to make make a bit sense of this. What we can see here is really a demo agent. It’s a demo agent customers deploying this production like do this differently.

But the the idea here is that you have now frontier models u be you know connected to your simulation software in a way where context construction prompt engineering guard rails data provisioning etc is done in the background. And so what we now prompted the agent here is say can you evaluate the peak stress of this model use the cat labels and can you see the last sentence?

I also prompted it that this is a live demo. Please try really really hard not to screw this up. So let’s see how this goes. And what it now does, again, it’s a bit of a funny demo. So, customers deploy this differently, but think about it. This is now able to use physics. It can reason through physics. It can reason through topology. It can find certain CAD labels, yada yada yada.

A customer deploying this to production would do it differently. They would give it guard rails, etc. The internet is horrible here. I do think it’s the internet, but let’s see how far it goes. I hope maybe to to to wrap this part up. The everybody that is familiar with how simulation teams automate simulation workflows, now we’re in the material section, right? It starts applying materials. Everybody that’s familiar with how simulation teams automate simulation workflows are familiar with everybody have all of the teams have different labeling strategies modelbased systems definition strategies how they connect the the SPDM system to simulation workflows etc.

So again in a real world scenario a customer would have a certain way how they automate things and you can plug into this workflows they would use prompt engineering fine-tuning etc to actually put these agents on rails in the way they want them to right so please just take it as a bit of a funny demo to see the concept it would set up boundary conditions now in a second I guess let’s check at at the end of the presentation and move over good many applause for the live demo All right, boundary conditions, right?

We’re seeing boundary conditions. Okay, good. Let’s move over just to wrap this up. I hope even with all the microphone handling and the agent being a bit slow, I hope the concept of moving simulation data into a structured accessible way, deploying it broadly, having the frontier models be thoughtfully connected and then have the ability to train physics AI models as you go deeper into design spaces shows that this is not just a you know like an an AI methodology that’s supposed to sit in a lab where somebody you know has a very specific AI model only this person can use but we’re actually an engineering team can start leveraging AI practically thoughtful right I hope that that came came a bit across just some use cases am I still okay on time doing okay there is a we have a fairly large case study section on the Sims website so if you’re interested in how different teams are deploying this we’re working across industry we’re working with engineering organizations at many different maturity scales on the AI adoption so if you want to take a look there’s a large case study section where you can click around and see if there’s something useful for you.

A few examples on what you can now do with this or what we’ve done with it or what have customers done with it is one interesting aspect we’ve shown this at NVIDIA GTC a couple of months back. An interesting aspect about the deep learning surrogate models those model architectures generalize better and better. There’s more and more model architectures being proven out. There’s a lot of interesting research happening and so we put this to a proof or we put this to a test and trained a fairly broad a pre-trained pump CFD model that now out of the box can at least sort of estimate the CFD simulations of an unseen pump.

Customers typically use that then to do finetuning so they train more on top of it but what it then enables is instant almost instant design optimization. And so we can see here now a physics AI model being hooked up with a parametric optimization run. And so on the right this is about 300 designs 400 designs of pumps sweeping through a design space. Again parametric optimization is is an old concept.

What’s interesting here is this took 5 minutes right because again instant prediction and so it’s the the realm of where shape optimization can be used becomes bigger. One example customer that put this to use. For a specific example we can see a this is a leaf blower. So an appliance example that’s a public model that has been used for pre-training and then this is the customer model where the sort of the finetuning of the of the physics AI model has been done and then again the the the technique we’ve just seen to do a parametric optimization for this kind of actual fan which was an interesting one then some academia this is the correlation with with the actual simulation results we see engineering teams big engineering teams deploying this broader and broader ITW W big tier one from the US applying this technology for nonlinear FA of snap fits.

And then a couple of news around technologies. A year ago Simscale did not support electromagnetics simulations yet. The portfolio of low frequency electromagnetics becomes bigger and bigger. I’m a CFD person. I have only a vague idea what these things are, but hopefully you you know some of the people here have. And then with a new partnership with Hexagon, we’ve integrated Hexagon’s mark. People might be familiar with it.

One of the you know I would say one of the most advanced nonlinear solvers because we saw a big pull from the market. And so now topics like you know consumer packaged goods, ceilings, bushings highly nonlinear applications become possible in Simsll as well. That’s one of the news. And then just to conclude we have Brad and Max in the background. Max has been checking Facebook and I think also Brad is not paying attention.

19:10 So we have Brad and Max from Entop here. Now, Simskull also natively integrates with Entop since a couple of weeks. We launched this just in in LA with the team at NTOP. So, you can now bring in implicit geometry natively. You don’t need an STL. You don’t need a facetation. Can bring it in in Simcale. We also don’t facet it. We generate a background grid right on the implicit file.

19:32 People have been putting this to good use already. And then again maybe to conclude and to bring this home everything inside simc scale automatically can be set up aentically. We’ve seen the hexagon mark simulations for example they out of the box on the first release can be set up identically. You can now set up nTop simulations inside simcale identically and all of this works out of the box with physics method.

19:53 So we can see here on the lower left a small TPMS physics AI model we’ve trained to predict the safety performance of different TP TPMS structures with in in real time. Good. To wrap this up we’re personally very very excited and I and I think like in the lunch break when we talked about it I think everybody feels the same way that we think it’s so obvious now that this is coming.

20:18 This still needs to be defined. So there’s a lot of dust that first has to settle but no doubt engineering will look different in a couple of years. We think we see engineering leaders understanding it. Engineering leaders are actively trying to figure it out and I think it’s an exciting time sort of this community is trying to figure it out. We’re excited to be part of it and I’m here today all day.

20:38 So if you want to chat further on that, I’m happy to. To learn more about the CDFM computational design symposium series, to see the archives of previous presentations, and to learn about future events, visit CDFAM.com.