0:26 All right, hello everyone. I’m Andrew Acuff. I am implementation executive at SimScale customer facing. I do not have a technical background. I actually come from mental health where most of my early career was trying to bridge the accessibility gap by training practitioners to work in family care practices, primary care models, which is interesting cuz it’s not that dissimilar from what I’m doing today where we’re trying to move decisions at the point of need and make these tools more broadly accessible.

0:57 We can design some really amazing technology, but can we deploy it at scale and to allow our our our customers to be very productive with these tools. So the talk is going to be moving from months to minutes and how AI is accelerating the simulation driven product design. I pivoted on the name or the topic. If you look at the agenda, it says reshaping, but I don’t think we’re really reshaping.

1:30 We’re accelerating. There’s some fundamental practices of engineering that we want to retain and I’ve heard this over and over about increasing iteration cycle, keeping in engineers in the loop at times. So what we’re trying to do is just remove the unnecessary wasted time and truncate this process. SimScale for those who are not familiar, we are an AI native engineering simulation a in the cloud accessible from a web browser, broad physics.

2:00 Again, what we’re trying to do is make tools more accessible broadly, and kind of take them from that sort of analyst realm and allow more engineers to get their hands on them. So, what we’re going to talk about today, we’re going to look at the current state of AI in engineering, just set the stage for what problems we’re trying to solve, and how we’re going to solve those problems with SimScale.

2:23 And a little bit of a spoiler, it’ll be physics AI and engineering AI, where we can create predictive physics models for instantaneous results, and then leverage engineering AI to accelerate a lot of the manual work with purpose-built agents. And those familiar with SimScale, we think that the cloud really enables these technologies by allowing broad access to advanced compute, organized data, and central governance. We hear that time and time again about having the data available to train, so having all that in one place, where you can easily deploy or easily train these models to be deployed.

3:06 Starting in 2025, we we started this state of engineering annual report. And in frankly, in 2025, it was pretty underwhelming. There was a lot of buzz, but there wasn’t a lot of use cases, there wasn’t a lot of adoption, there wasn’t a lot of value being produced. We did that survey again this year, and it was very, very different. We’re finding that these tools are producing a tremendous amount of value, and this was across a range of industries in the the US and the the EU, and found that, you know, more design variants are being explored, faster RFQs, faster simulation turn times, and really broad adoption, just depending on where companies or engineering teams are in that that that process.

3:50 So, just drill down a little here. First finding is AI is expanding the design space. We found that there’s three times as many design variants that are explored using AI tools. So, a a broader range of design variants. But, if you drill down, if you look at the conventional workflow, there’s only about 3% that we’re doing more than 50 simulation iterations on a design. Whereas, if you look up at the AI workflow, it’s over 30%.

4:20 So, a huge delta between the conventional and the AI. AI is creating beginning to create some some significant commercial value. If you think of three times faster design cycles, if you think of three times faster RFQ responses, that produces a tremendous amount of commercial value for the companies that are adopting these tools. And again, on some further drill down, if you look, there is 0% of conventional workflows that are getting these done in less than an hour.

4:58 And only about 1% that are doing that in 1 to 3 hours, from our survey at least. Whereas, the AI-enabled workflows are 40x that. So, a significant delta again between the adopters and those who are sort of the laggards in that space. And then, I don’t think this is news to anyone in here. You know, AI is being adopted. For engineering teams, I don’t think it’s a question of of if, it’s just when and how.

5:27 It’s who will adopt these and be successful, and and when that will be. But, if you look at it, you know, mature AI programs are less than 10%, and there’s about 80% that are working in pilots. So, it is being adopted, but we’re still very very early in this in this cycle. So, thanks for all the work everybody is doing. So the problem we’re looking to solve with AI it’s it’s essentially been solved in software.

5:58 So, you know, major companies are deploying thousands of lines of new code a day. It’s it’s a very fast cycle for software, but sim or hardware still takes weeks. And we believe quite firmly that you know, AI’s biggest impact will be on the physical world. That’s just still yet to come. And why is that? What what kind of brought us here? So, you know, why is that?

6:27 Why is the challenge for the hardware side versus software? Host of reasons, but if we look at the the traditional tool stack of simulation engineer, you know, maybe they have one code or solver for a certain physics. You know, they’re working on a workstation maybe under their under their desk. You know, now they’ve got maybe a pre process preprocessor for another tool. Maybe they’re in a cluster.

6:54 You know, they’re writing some rigid scripts that if you know, one little dot gets off, it’s going to crash on them. Maybe they’ve got some CAD integrated tools. Maybe they’ve moved to the cloud. Now you got your engine or your design engineers that are very often times in a different tool stack. And we we sort of made this slide into a cluster cuz it’s kind of what it is.

7:16 A colloquial term where I’m from is is a rat’s nest, right? That’s just a it’s just a mess of tools and software. And we’re trying to streamline that. And so, they’re siloed. They’re typically manual. They’re slow. They’re running on local resources many times. So, now you start aggregating that weeks of a design cycle on one part. Now you think of a car. Now, you know, you quickly get to to years and years.

7:40 So, we’re trying to shorten that. But, this world is very difficult difficult for AI to enter because it’s so fractured. You know, I think what you know, there’s this expectation that we’ll get to this point where you just say the end point you want and it’s just going to spit it out to you. That may come someday, but we’re a ways away from that. And just because of the way that these tools the data is not there or it’s fractured, it’s hard to train.

8:13 The tools often times don’t communicate with each other. So, this is just not a reality today. So, aside from the tools, let’s look at the actual design cycle and and see where a lot of the time get lost. So, say we have a a design engineer on the left. They’re sending work over to a similar an analyst on the team. And just a a hypothetical, let’s say it’s a manifold here.

8:40 Well, you know, what are the ways that we can most dramatically improve that process? Well, you know, something we’ve heard time and time again, I’m really excited to hear some of the talks here around quantum computing. Maybe quantum computing holds that promise for us. Maybe it’s a lot of GPU accelerated design, but there’s trade-offs. There’s no perfect solution. Everything’s trade. There’s going to be economic cost to these.

9:05 There’s going to be a lot of new code that has to be written. For us, we think the immediate solution to the speed is physics AI where you can train a model on predictive physics and then use that for instantaneous results. So, you kind of front-load the computational cost and time and then now you got a this really efficient model to use time and time again and and continue to iterate and refine on it over time.

9:35 But, where we think the biggest time, suck is, where where it’s really being wasted is if you notice the gap on the simulation lead time, this back and forth between teams, handing off from one department to another. You know, you get a backlog, an analyst has got a backlog of projects. Maybe it’s not a huge priority for them. Maybe they want to work on a a much more, you know, or have a priority for a much harder problem.

9:58 They don’t want to focus on some little bracket, right? So, workflows are really bottlenecked by a limitation of simulation experts, and just this back and forth time. So, we’re seeing, you know, an automotive part, maybe a couple of weeks. Aerospace could be as much as a month. And then, you know, in electronics, we’re hearing customers that are that are taking 8 weeks of this process before they can nail down a design.

So, you know, there’s a lot of strides to be made in the solve time, but fundamentally, it’s a workflow problem. So, we’re trying to solve this workflow problem. And we do that with engineering AI. So, deploying agents to accelerate a lot of the manual work, cut down a lot of this back and forth, or automate that that process. So, the solution, keeping what works, eliminating what doesn’t.

10:52 I’ve heard this from several speakers, the value of this iterative process, right? So, let’s take this manifold design again, where, you know, design, simulate, design, simulate, this back and forth. We’re not trying to get rid of that. What we want to do is just really speed that up through engineering AI, where agents can drive the design iterations autonomously, physics AI models, which will allow the the AI surrogates to produce results in seconds.

11:18 And then, of course, this being facilitated by cloud cloud resources, so you have organized data, you have compute elastic compute on demand, and you can kind of get the your organizations at scale onto a similar page. Zero shooting as a reliable design that may come someday. We just don’t think that’s that’s really here today. So not to be disparaging against with anybody working on that. We just think the most practical solution for that today is iterate or AI systems that can iterate through thousands of engineering design decisions in hours.

11:59 So again, might change the talk from reshaping to accelerating. There’s fundamental utility in some of these these workflows. They’re just taking too long. So how do we accelerate that? How do we truncate that? So how can AI provide a solution on the engineering AI side? Natural language so so user intent driven moving away from rigid scripts to natural language models that understand context, user intent, can be custom built, custom trained for specific specific use cases that allows it to plan, reason, execute, set up workflows, execute workflows autonomously.

12:50 It’s able to operate the platform. You know, handle uncertainty. You know, we’ve got under the hood we’re we’re we’re leveraging Claude because that’s kind of the best in class today, but it’s an open platform. If another solution comes along, we’re happy to we’re easily migrate over to that. So let’s see this in practice. So just an agent which is prompted to set up a simulation. Goes through as as some you may know there’s there’s a A of time wasted in CAD cleanup, CAD modification, especially if you’re let’s say an analyst and someone sends you a production ready model that you need to do a lot of cleanup on.

13:35 So, going through the the agent knows to to do some defeature to simplify the model. So, it goes through a CAD editing function here. And if you notice it’s of course all documented there. So, sort of a a way of knowing what’s this traceability and and what it’s working on. And once it’s got some updates to the CAD, some cleanup, and I think this is running Yeah, I think this is What is this running?

14:14 I think this is running like a vibration analysis, structural analysis of some sort. Okay, so after it’s got the geometry cleaned up, it’s it’s sort of a simulation ready. It’ll ask and now of course you can prompt it for a specific use case, but it it infers or it essentially guesses what what the next steps would be. Again, this is all documented for you. So, it’s what’s the I guess what’s what’s the goal of of the potential agent?

14:46 Is it to allow more novice users, people that are you know, engineers that are less familiar with simulation to run simulations that they traditionally have not been able to? Or is this a way for a more experienced user to offload a lot of manual tasks, a lot of things that take up their time that are not necessarily high value? So, it frees them up for higher decision making, more high value times of activity.

15:12 And then towards the end, once it’s got the analysis ran, I think it asked if if we want to generate a PDF. And that PDF is produced. So again, no one really likes writing reports. That’s and that’s just one example. You know, here’s some other ones, you know, run up thermal vibration and stress you can do you know, some CAE automation, whether that’s design exploration, optimization studies, code compliance.

15:48 You know, this can be kicked off from Slack or other channels. I know the you know, just the idea that one day you’ll be able to or you could be here in a conference and just from your Slack kick off an analysis. Maybe someone on your team ask you a question on something, you could just Slack it and have a an answer for you during the talk.

And it’s we actually I borrowed this from our partners at Buehler. They gave a talk here recently and it’s just a way of visualizing this this sort of sh- bit of a shift in the workflow where you know, the the knowledge is concentrated into a few specialists and now they’re backlogged. There’s a this iterative back and forth. Whereas you can train these models, whether that’s, you know, a a surrogate physics model.

16:36 And some ways you’re creating sort of a surrogate of your your engineering expertise as well and codifying it so it can be scaled to the team. Now for the physics AI side, how how can the solution provide there? So phase one, train. Phase two, inference. You know, set up a you know, a parametric study and kick off a a lot of solves in parallel, train your physics AI model.

17:13 Now it can be deployed to the team. So you again, you front-loaded that all the solving and so you can get instantaneous results in the future. Very similar, we’re we really focus on the platform. You know, SimScale started building a GUI on top of OpenFOAM, Code Aster, some open source code. And what we found is over time we we identified capability gaps. So, we’ll go out and license other commercial solvers such as MARC.

17:42 You know, we got a lot of spells, but what we focus on is really the deployment and the platform. And then we just plug in what’s needed. So, same thing with our our learning models. You know, we didn’t choose a particular horse there. Different models work for different use cases differently. So, we have access to to various models depending on the use case. And here’s just an example in the in the step change.

18:08 So, think, you know, rewind time to only analysts were driving simulation tools. You got design engineer that needs a pump. They send the model over. The analyst, maybe they look at a couple of flow rates. Then they send it back. Maybe there’s a couple of design iterations that need to take place. So, this back and forth maybe took months. Well, fast forward a little bit and you start getting you know, easier to use solid analysis tools embedded in your CAD software.

18:42 So, now you kind of move that upstream a little bit and engineers are running some of their own simulations, but they typically run it on local workstations, unlimited computes. Maybe there’s licensing restrictions. They can only run on so many cores. They can’t run in parallel. So, that’s still, you know, even let’s just say you’re looking at a performance curve and you need 10 flow rates, 20 flow, whatever it is.

19:06 That still takes weeks because they’re having to and it’s tying up their workstation, right? Cuz they’re all being run serially. Well, then comes along cloud simulation and we can truncate that down into days. Now, instead of running them serially, set up a parametric study and run them in parallel. Breaks it up on a bunch of different machines, get results in an hour or so. So, just from there we’ve gone months, weeks, days, and now with a predictive physics AI model, we can get down to minutes.

19:33 Okay, so these are these are the step changes that we’re seeing in the in the solve. And Arman will explain this much more eloquently than I would, so I’ll certainly recommend anyone to to look at the webinar that we did with our partners at Conneon. And this is exactly what they did. So, they they trained a physics AI model, ran a large-scale DOE study, ran hundreds of simulations in parallel to map out the full space.

20:00 Now, whenever they’re looking at a new design iteration or a new project, they can get these predictive results you know, or do a full study here in about an hour. And of course, we certainly always, you know, the same way you would want someone to sort of benchmark an analysis, you want to occasionally PD solve one of your AI predictions. So, it’s just constantly fact-checking, circling back and fact-checking reality.

20:31 Just doing it in a very efficient, cost-effective way. So, let’s put this all together and just kind of look at how these two systems can work in parallel. So, we got a a server rack here. I think we set this up, maybe 20 different fins for for both of the different heat sinks, a handful of different positions for the fans that have been sort of idealized there.

20:56 Of course, this is sped up, but just showing that again, the the setup is fully automated. So, it’s running through a setup, it’s going to go ahead and kick off a simulation here in a moment. Signing material properties, parts. And again, yes, this is sped up. This would not occur this fast, but it would be occurring in the cloud. So, get right back into whatever you’re doing in your tool.

21:28 It’s not tying up a workstation. We actually had one of our engineers who was giving a talk here recently and was running simulations on the flight over, which is I thought was pretty interesting that that’s possible today. And so, now we can look at the temps. You can get lunch right on time. You know, put some point point probes down, check your temperatures. Once the analysis is finished, again, this is all in cur occurring in web browser.

22:04 On the physics AI side, you can see our little magic wand up there, if you will, that will kick off the the the results here. That’s actually in lab time to get those results in seconds, and we typically like to calibrate that within about 5% of accuracy from the the solve. Now, just drag in a new geometry, if you want to test it out. Hit go, and again, results in seconds.

22:30 So, now, putting these two technologies together, you can automate that iterative process while getting instantaneous results. So, you can run through this study, which, you know, again, would have taken months potentially and and get it here in just a matter of minutes. And that’s all I got. Look forward to meeting everybody here.