0:15 Good afternoon. I’m Wasil Rezk, the CCO of BeyondMath. We’re a Londonbased AI company building foundational models of physics. And I want to start off by saying that, you know, we are designing things today that are order of magnitude more complex than they were 30 years ago. More physics, more multi-disiplinary coupling, more design variables. But the simulation tools that we have today are still running the same way they did decades ago.

0:43 And I think you already know that we are writing a mesh, submitting a job, waiting hours, we get an answer. But the complexity has scaled a lot more than that. Has scaled exponentially. And the tools haven’t. So, we’ve built something that hopefully closes that gap and I’m going to show you throughout this presentation what it looked like. So, to give you a sense of the scale of the problem, a few hundred milliseconds of real-time physics can take a $50 million supercomput many hours, sometimes days to compute.

1:19 Running these tools properly requires a PhD in physics as you know or an expensive set of consultants. Most design decisions are made without many simulations because teams can’t simply wait or afford to wait. And in that because of that engineers are simulating a handful of designs when they could be exploring thousands of designs and exploring an entire design space. And as a result what ends up happening that performance gets left on the table and we don’t even quantify how much is being left on the table.

1:56 So at BeyondMath what we’ve done is we’ve built a foundational AI model of physics that’s delivered through our platform GP studio generative physics studio. So let me just be precise about what it means, what that means and how it differs from surrogate models which I’m sure you’ve heard about. Our models is trained on the underlying physics on first principles data that we generate ourselves. It learns the governing equations, the transient dynamics, the causal and the causal relationships.

2:26 That foundational understanding allows the model to generalize across unseen designs. Hence removing the need to be continuously retraining your model which you find in the surrogate approach. We do ye we do use CFD simulation data but an order of magnitude less than what a surrogate model might require. So when we work with a customer, they might share with us data that they have already, right? It’s data that they’ve invested in, but the amount required is a lot less, an order of magnitude less.

2:59 And we can I’ll give you examples of what that means. And it the model can generalize to new designs. So once you once we’ve built that model, we don’t need to go on and continue retraining it, you know, if they want to change something dramatic about the design. A perfect example is if you’re designing cars and you’ve trained a surrogate model on you know a set of cars and you want to add a spoiler.

3:23 Guess what? You need to train more more geometries and more data to include that in the target model. In this case, you wouldn’t need to you know make sure that you have spoiler spoiler in your geometries. It can generalize to something like that. So three things make our approach quite distinct. The first is that this is a new category obviously. So we are the first mover shifting from the static surrogate models to an architecture with a generalized understanding of physics.

4:00 So it’s not incremental. It’s a completely different paradigm. The second is that we train on transient modality not stat. So the model learns the why and the how of physical movement not just the end state. It understands the temporal chain of cause and effect and that’s what separates learning physics from memorizing results. And the third because the model understands the fundamental laws of physics it generalizes to novel designs it has never seen.

4:25 As I said before the surrogate model needs retraining when the geome geometry changes significantly and in our case it doesn’t. So the core difference across the board is is between learning data correlations and learning the actual physics. Surrogates as you know are trained on large volume of customer simulation data. They’re fast at runtime but they’re quite brittle when it comes to extending outside of that training data.

4:50 In our case we need an order of magnitude less data. We run in seconds and we generalize across unseen geometries. So I’m going to show you what that looks like in practice. So we started off actually in in F1. It was a natural proving ground press proving proving ground for us. From there we moved on to other fields automotive, aerospace and electronics and I have examples from that as well.

5:21 But in F1, as you know, it’s one of the most demanding engineering environments you can have. And the teams, you know, on average spend around $3 million on their CFD simulations. They have a lot of restrictions on what they can do with that, obviously. But what is really crazy about F1 is that they are making changes to the car before every single race. And every track has different characteristics.

5:47 So what they do is they actually run a air map which is a matrix of simulations across the different yaw angles, right heights, all the kind of potentially different positions across the the track and then based on that they will find the optimal arrow setup for that particular race. So today because of the compute constraints they have and just in general how slow the traditional solvers can be they’re running about five simulations per week.

6:19 So more or less that is the amount that they have before going into a race and that’s they have to infer all the changes they need to make based on those simulations. And we are working already with two two F1 teams where they’re making they’re running thousands of of sims on our platform. And what that enables them to do is actually make much more well-informed decisions when it comes to the changes they’re making before every race.

6:47 And I’m going to show you here. Hopefully it works. Yep. So what you’re seeing on the screen is so this is G from GP studio our platform and it’s running a full 500 million cell simulation of an F1 car. This is not an belong doesn’t doesn’t belong to any particular team. It’s from the FIA the governing body. And the key point here is that the model has never seen this juncture before in its training data.

7:16 So this is a generalization clearly in action and the model has been trained on again transient sims right of different of different F1 sim data. It’s also been trained on essentially navaric equations the underlying physics for aerodynamics and you can see that it’s producing accurate results. In fact, I I believe this is within 2% error score of a full DEZ CFD run. And you can see all obviously all the wave structure coming out of the the rear wing here.

7:59 I’m going to show you another example in automotive. So we moved from F1 as a company to as a natural step into automotive. We work with mainly kind of automotive OEMs in this space and here their range is a bit more wider. They have a bit more flexibility obviously in the way they run simulations. They don’t have a governing body, you know, on top of them telling them what they can do.

8:23 And the main thing they’re looking for is obviously reducing the cost of HPC simulations, right? It’s it’s it’s it’s a lot. The big line item in their budgets, but actually how can they obviously increase unit sales and that all goes down to range, right? And and drag reduction. And and that’s for, you know, your average major OEM. That’s the main that’s the the main goal. And in our engagements with with these OEMs, we’ve been delivering predictions within 5% of the full CFD.

8:59 Again within the acceptable accuracy metrics that they require. And again, just to clarify, when we’re working with these teams, we’re, you know, we’re building these models with an order of magnitude less data than what a target company would ask for. And what that means for an automotive OEM is that they’re not having to invest in hundreds of thousands or millions of dollars upfront in generating that training data and having to continuously manage that training data and continuously add more to it as the the design space widens for them.

9:37 So I have here an example. So this is GP Studio again in action. And what it will be starting to do is run drive air geometries. So drive air geometries is a open source benchmark in the automotive industry. It’s it’s widely known and what this demo shows is our model predicting the full external flow fields around the vehicle body. And it will start to also take in different variants.

10:03 So these are all variants that are already in the open in the open source in open benchmark and you can start to quickly iterate look at compare different results and again the model here has not been trained on the entire data set. It’s been trained on a very very small section of it and we’re planning to actually release kind of our white paper or our paper for this very soon.

10:33 Another benchmark in automotive is is air SUV. So this is what you you’re seeing here. So this is again a sideby-side comparison of of our model compared to a open foam CFD solver running you know I think dezzim here and this you’re seeing a similar thing here where again we’re able to match the level of fidelity of the 143 million cells this is pressure gradient across and with very high levels of accuracy again usually typically within 5% % accuracy here 5% error of a full CFD run and the same the same conversation here we’re training on a very very small section of the training data available and in terms of performance we’re running in seconds compared to the the CFD solver and this only c only costs a couple of dollars to run so for one U and actually aerospace is really where the economics starts to get really drastic in terms of the scale that we’re talking about.

11:51 So we’re part of a consortium with a tier one supplier Honeywell in the UK to redesign the aircraft components that they sell onto they manufacture and sell onto Airbus and Boeing. These are all kind of internal aircraft components and here it’s very similar. We’re providing kind of the simulation AI simulation and optimization tool part of it. And if we could just increase the efficiency of those designs, make them more lightweight, that translates to $2.5 billion in annual efficiency saving the flowing and Airbus.

12:27 U we just started this project in September 2025. So, we don’t have any results to show, but hopefully very soon we you’ll start to see some stuff publicly. And it’s not just on kind of the large scale engineering or or kind of just general aerodynamics. We’ve been actually applying our model when it comes to thermal management and in this particular example on the Raspberry Pi Pico device.

12:57 And what you’re seeing here is the heat distribution across the components, the thermal hotspots and how the heat decipates throughout the device. And a similar thing here on GP Studio, you have an LM interface. You can easily interact, visualize your results, run these sim simulations in a couple of seconds, compare the different variants and and run more based on run more simulations based on, you know, the direction you want to take the design.

13:32 And again obviously with thermal simulation it’s one of the biggest bottlenecks today we see in electronics you know and more widely in you know in a in a very different industry but quite similar in in when it comes to thermal on you know semiconductors and our goal here is if we can allow an engineer to assess hundreds or thousands of designs quite quickly start to to run these sims right quite fast with very little upfront investment And we’re unlocking real innovation here, right?

14:04 This is being able to use an actual AI model and actually have the model working for you rather than the you working for the model. I think that’s the biggest critique that we we are seeing with the surrogate approach and we believe this is a a viable alternative. So everything I’ve shown you here is about how our model is predicting physics on existing geometries. And I think there was a previous presentation talking about it’s great if you have an amazing foundational physics model but if you don’t have the ability to generate enough designs then the model the physics model kind of becomes useless and I and I agree with that I think we with a lot of our customers we see that you know they appreciate this technology but they want to have the ability to generate hundreds or thousands of variants or just new designs a lot quicker so it’s something that we’ve been working on.

15:01 We’ve kind of called it a dantic pad. We might change the name. Might I don’t like the name as much. But essentially it’s a design tool that understands the physics natively. So what I mean by that is if we have now the capability of having a foundational AI model that’s really useful across F1 or automotive or aerospace we can if we work on a CAD tool that basically can generate parameters generate all the different components but has the physics baked in as in every design it’s outputting is physically viable right with some cons with some constraints that you can predefine if you want to predefine then that is extremely powerful.

15:43 So we it’s a complete paradigm shift where instead of doing designs and then moving on to simulation, you’re just focusing on the design. The simulation is h essentially the simulation is happening in the background. And we we are we’re taking a bet on this kind of paradigm shift. We think that in the future you don’t need to even be thinking about the physics. The physics is baked in.

16:09 And we’re starting to explore more and more of of what that means and how that looks like obviously across different industries. So I have here an example and this is just a demo that we came up with internally is a propeller where it’s we’ve generated it automatically from our pool. And we’ve given it you’ll see in the prompt kind of very little instruction. We’ve kind of said, hey, just generate a central hub of 12 blades should be twisted blades with the best height blades around it.

16:42 And I think there’s so you’ll see here in the response it kind of what it starts doing is it start generating the different components. So you can see on top on the top right there’s some topography of the components and you can auto automatically generate parameter. You can play around with these parameters. The idea here would be that instead of instead of now running a physics simulation that is already done in the background, you can check it and you can verify it.

17:11 You can validate it with CFB if that’s important to you. But imagine a world where you are iterating on your design. You’re not having to worry about how viable it is physically. It’s all being kind of taken care of. So that is the world that we’re moving towards. We’re already testing this out with some existing customers we have. But we really believe in a future where you don’t need to be iterating continuously between a design and a simulation team.

17:45 So I open by saying that kind of complexity of what we design has of what we design today has outpaced the tools we use to validate it. And that gap is not going to close anytime soon and it’s not going to close on its own. I think it needs a fundamentally different approach to physics simulations and the way we even think about the way we conceive how we are building things and I really I think before in a in a previous presentation it was kind of it resonated with me of like you know we’re taking legacy ways of you know a legacy design of a boat and we’re just that’s why we have the boat we have today.

18:21 I also think we shouldn’t think from a legacy perspective when it comes to the way we build things. The actual process of how we even conceive or design something. And that’s the direction we’re heading. We think physics should be baked in. If you have a foundational physics model that works and is giving you the levels of accuracy required, it actually all goes back to the design. The design is where the kind of creation comes in.

18:48 And that’s where it matters. So, thank you for listening and yeah, hope to see you. To learn more about the CDFAM computational design symposium, access the archive of previous presentations, interviews with speakers, and information about future events around the world, visit CDFAM.com. Come.