All right. Hi everyone. My name is Mike. I lead research at, Arena Physica. I’m super excited to be here. Share a little bit about what we do on art artificial intuition. We’re focusing on building electromagnetic magnetic solutions. Particularly on the modeling front. That’s what I’ll be talking most about today. Maybe just to give you a brief in introduction about us, Arena Physicica, our main platform is Atlas.

0:49 It’s an agentic platform very broad in integrations very deep on context for hardware systems plus software systems and then essentially focused with proprietary tools and and solutions that we build on top of that for specific workflows. One of our strengths particularly is in the EM field or in particular where we also have a lot of talent both in the industry but also in academia. The second piece is then labs and hardware services.

1:20 So we use that also internally to deliver those services and even build hardware. And then the last but definitely not least is heavyside foundation model effort and really excited to talk to you guys and bring sort of the latest of that effort to you today. We deploy Atlas across multiple different customers. You see some of them there anywhere from you know cluster deployments GPU tuning to sort of the defense space and also the a bit more pharma.

1:48 So quite a flexible platform hardware systems is the specialty of it. It’s actually quite an exciting time for hardware innovation. I’m preaching to the choir to some degree here. It’s really great to to be talking here. Seeing a lot of the struggles and a lot of the pain points identified that we definitely also share. So when I’m talking about hardware innovation, what I’m really excited about is the pure ambition that we have today about what we’re going to solve and how we’re going to solve some of those things.

2:15 For example, we were talking about data centers in space, energy harvesting in space. The fusion reactor in France is scheduled to go live within the next decade as an example. We have quantum computing, we have autonomous systems. All of these things are big dreams. In many situation though that hardware innovation actually means novel designs and I mean truly novel designs truly novel applications of the physics that underline those designs not design iterations or just adaptations of existing designs in some of these cases.

2:48 That’s what we’re particularly interested about and excited about. So if we knew if we know we need this innovation and specifically in the EM field the challenges are actually massive. If you looked at for example a car a little while ago let’s take it 10 years you compare the electrical system what it is versus today the the difference is shocking right and that persists through all the systems pretty much that you know we all know the problems of smaller dimensions right that come with our compute requirements those small dimensions by themselves are difficult enough if you’re talking about EM and higher and higher frequency requirements they’re catastrophic in some senses your entire designs change.

3:30 Those designs are fully 3D sometimes. So the the challenges there are quite massive. If we go back just as the last one to our space example, now we also want to put all of that in space. So things like radiation hardening where all of those things become even more important, not all of which has solutions as of today. I have one example on the right hand side.

3:49 I’m going to use this quite a bit for the rest of the talk just as an example. Some of the things we’re striving to do. This is a semiconductor packaging case just very schematical in the top from individual sort of dice on an interposer with connectors on a substrate with connectors on ultimately a PCB with connectors. All of those things are electrical connections that we have to simulate.

4:13 They’re at ultra high frequencies tolerances are small and tight. So those are the type of problems that we’re really interested in. You can see a scanning tunneling microscope image of that just as a as a really nice visual. But even for when we jump back to EM now even if we’re talking about simplified 2D 2 and a halfD structures the design process is quite difficult and in the last talk I gave in Barcelona I was talk I was calling it dark art deeply on for almost all of us and and actually a true innovation for for or a true bottleneck for innovation I should say.

4:51 I’m going to give an example here of a structure I’m going to use as an example for the rest of the talk a little bit. This is a typical human-designed RF filter structure, right? You can see it’s essentially planer. Yes, it is quasiplaner. It’s layered. There’s a ground plane on the bottom. There’s some connectors, but it’s ultimately a 2D structure. And on the left hand side, you sort of see the metal structure on top.

5:12 You see a port on the bottom, a port on top. The port gets excited. At 7 and a half gigahertz the electromagnetic field does not penetrate from one end to the next acting as a filter or a blocker and then at 12 and a half for example that same structure is is actually transmitting your signals. So that’s a bit the the type of setup that we’re talking about the type of structures we’re really excited about simulating all of this in 2D for now or 2 and 1/2D for now.

5:41 So if all of this is really difficult and even designing these in 2D or is is still today a dark art difficult to do what’s a path to actually an AI assisted DM design that is better right and so for us u understanding of the underlying physics I really appreciated the earlier talk the importance of grounding all of this in actual physics not a language description of part of the physics that we may have already done at some point but the physics itself that’s certainly something that we believe is truly needed.

6:13 An efficient generation of valid designs. We heard this also a lot today already, right? If we can generate more and more valid designs quickly, cheaply, thousands of them at a time, that has value. And then lastly, how do we actually evaluate them cheaply and fast? Again, particularly in the EM space, that is not a given. Anybody who’s worked with these simulator knows they’re extremely difficult to set up.

6:37 First and foremost first the meshing part, then the configuration of the simulator themselves, then all the boundary conditions around it. And even if you do all of that right, you may still actually end up spending one 2 3 days on relatively simple geometries and simulations. So that’s sort of like where we are. That’s the motivation of why we want to tackle some of these. So how are we tackling this?

7:03 At Arena Physica, what we’re building is a foundation model for EM. And I want to just say quickly for a second what we believe is a foundation model. Foundation model is something that learns the underlying physics, not a mapping for one application. It’s something that learns a variety of input spaces. For example, not just a very narrow frequency domain, but maybe a large frequency domain, maybe at some point, ideally all the way up to optical applications.

7:34 Right? So that’s what we mean really a broad set of features that we can then use and some of the stuff like sparameters fields radiations gain factors etc can be then calculated from those underlying physical predictions that you make. We actually built two models heavy side that’s our foundation model. It’s a forward model. We just had a bit of an introduction of forward and inverse design a little bit earlier.

8:00 That’s our forward model. It takes you from a geometry, materials in space arranged a certain way, excitations, boundary conditions to EM behavior. On the flip side, we have an inverse design model as maronei that takes you from a desired electromagnetic behavior such as an parameter, a field pattern, a radiation pattern of sorts to a valid geometry. So those are the way the two things we actually build them.

8:26 I’m going to spend just a couple minutes of telling you a little bit about Heavyside. Heavyside is essentially a transformer-based model. It’s akin to what you would know as an LLM, right? But instead of the tokens being languages, the input tokens here are geometries. So materials, ports, via placement, thicknesses of materials, material properties, etc. And the output tokens can be things like fields as parameters other derivative of essentially the the physics that you care about.

8:56 Our first version worked essentially as you see below. It took these 2D structures and simulated what most people would care about for these specific 2D structures which are these scattering plots. Right? So the answer essentially a question at a given frequency how much gets reflected how much actually gets transmitted. Right? And so you can see an example of this on the left hand side again similar filter structure that I’ve already shown previously.

9:24 And then on the right hand side you can see the magnitude of these plots over frequency space for a heavyside prediction for the same structure an EM solver traditional sort of classical commercial solver and then also a test in lab. The inverse model mone to also give you a little bit of an idea is akin to the image generation models you might be using today. So diffusion process that really allows us to constrain condition for things like materials that we care about.

9:56 Maybe you have a certain design archetype you would like to use. We can condition for that as well. And then iteratively seed and den noiseise to actually get valid design candidates. Right? I tell you a little bit more about this. One thing I want to highlight about both of these models is here we’re not even talking about 10,000 simulations a day or 10,000 simulations in a few days.

10:15 We’re talking about 10,000 simulations in a second, right? That’s what we’re really after ultimately want to solve. So how is all this trained? We heard earlier training is super important. Data is scarce. Data is difficult to get. It’s expensive to get. The way we do it, we rely on the classical solvers that we do have. Of course, we use a multitude of them. We have our own.

10:40 We have a lot of experts in house. We start with expert-g guided procedurally generated geometries. Those are really targeted to actually just tease out the physics that we truly care about. Right? So think about coupling between two metal parts for example, right? So we can engineer some of these to do exactly what we want and sort of gear the system to learn the interesting physics that we want.

11:04 The second one are random procedurally generated geometries. We also do a combination of the two one and two just to learn anything. So even if it might not result in a useful antenna for example it still teaches a system a lot about physics right and then lastly physics informed there are laws there are theorems there are symmetries and I think if you want to build a system that obeys those you can use them quite well not only in the modeling side but also in the data generation side.

11:37 Ultimately though, one of the big advantages that we also have with these things, we can measure some of these things. And so we have our own lab where we closed one the sim toreal gap. So we can actually say what I’m not interested in is the pure physics that happen to be represented in a simulator today. But I actually also care about other factors that have not been represented in a in a simulator today.

12:02 Maybe we’re talking about a a specific machine, a specific manufacturing process that leads to deviations. That’s where systems like these are of course really good. You could do an a final fine-tuning layer for example for these. That’s one. And the second reason why we also rely heavily on these measurements is it actually validates the data generation at large. Right? Everything I’ve showed you so far is actually already working.

We’ve released a sandbox at least for people to play around with almost three months ago. I’ll summarize quickly because I really want to move on to even more exciting topics. But you’ll see on the left hand side for example how quick simulation has become the forward model, the characterization of a design. On the left hand side you see somebody drawing in almost real time. It’s a tiny little bit a bit sped up but not much.

12:56 Somebody’s literally just adding metal to a design and on the left hand side you see the target characteristics that you have update essentially instantly. Right? So when we’re talking about transforming design processes doesn’t get much better than you know removing the simulate button altogether. Right? In the middle you see the forward model and the inverse model in tandem. You see here hundreds at least or tens shown but typically it’s thousands of parallel designs at once.

13:25 The evaluation of those designs versus the target by the forward model and then the selection of one of them as a design candidate. And then on the right hand side because we have an agentic platform you actually see this all wrapped in an end to-end loop. So you every single person in here can actually go in one of the hardest fields which is RF and go and design an antenna with nothing or not an antenna sorry these are actually focused on on filters.

13:54 You can actually go and design a filter element with nothing than a few words quite exciting already as it is. What’s even more exciting we can actually show that this approach leads to novel designs. Right? So this is an example of something that we produced. We had a characteristic filter behavior from a known design and we told it can you reproduce something and actually came up with this design which every human would probably struggle to actually create right there’s no intuition left here whatsoever but it does work right it shows the characteristic behavior that you actually want and while this example is interesting by itself it does just show the path right what is what what are we capable of I think in terms of exploring ing new design spaces when we have a really powerful forward model and a really powerful inverse model.

14:52 But something’s still missing, right? If I go back to what we started, where we started, all those designs and examples I showed, particularly the semiconductor packaging space, how can we model some of those more complicated geometries, right? They’re truly 3D. There’s definitely nothing planer in those solder bumps or in those ball grids. How can we inform design choices better? I showed you scattering parameter, right? But if you would ask, okay, I know the scattering parameter is bad, but what am I going to do about it in my design?

15:26 That would probably be a very hard thing to do. So, we need a better way to actually have this information of what leads to good downstream characteristics that we care about. And I’m actually really pumped to announce today for the first time two breakthroughs that we make in our made in our modeling efforts. One of them is we now do full 3D input and output capabilities on this foundation model.

15:49 And the other one is we predict the fundamental of physics underneath it. Everything goes through the fields themselves. So we really predict the solution, the fields directly from Maxwell’s equations. And I’ll show you a little bit of what that looks like. I’m showing a mesh here because what we actually do is not a true mesh in the sense of the simulators you’re used today, but it’s a good representation nonetheless.

16:12 Before we had these 2D geometries, now we have space. That space gets discretized somehow, encoded somehow in our model. Doesn’t matter anymore. There’s no more concept of thickness. There’s no more concept of layers. It’s just space. And of course, what that allows us is to move eventually away from these planer geometry into things that are truly 3D, right? And I don’t think we’ve ever explored what filter geometries are actually possible if we leverage 3D entirely.

16:42 And I think that’s what we’re really excited about trying out with these systems. Relatively shortly to the second piece the fields probably even more excited we went to from predicting mostly what you see on the left hand side so multiple parameters across the frequency space to actually predicting thousands right of vectors in 3D space the fundamental solution to Maxwell’s equations and everything else including the S parameters that we see so far are then downstream derivations of those fundamental units Right?

17:17 And even just this simple example gives you an answer of why we’re so excited about this. Here you could instantly tell me where you have some coupling where you have higher fields or lower fields. That’s exactly the kind of design information that you would want as an engineer to actually go and improve this. Right? And now not only the engineer will have it but also the underlying agents that we use, the models that we use to come up with these autonomously.

17:43 I’m going to give you one example. It’s the same structure that we’ve been looking at alto together. If you’re looking at electromagnetic fields, there are field lines here specifically. So that just means the line follows the vectors in space. On the left hand side is a commercial simulator. On the right hand side is our foundation model. It has never seen that geometry before. You can see not only does it actually accurately capture a lot of the field information, but more importantly also the phase information.

18:13 I’ll show one more of these. If you have both E and H or B, you can also calculate the energy flow. It’s a downstream derivation of those two fundamental predictions that we make. Again, commercial simulator on the left, foundation model prediction on the right. That means we not only do E well, we also do E age well. The whole system starts to be represented quite fundamentally well.

18:38 And now, so what? We’re definitely not going to stick with a 3D representation of originally planer materials. No, the goal is really to go into the application space that is truly 3D. I’m going to bring back the example I made at the beginning from packaging and then you can see on top a structure that is simplified, yes, but nonetheless quite actually captures some of the physics that you truly care about, right?

19:02 You have multiple solder balls. They are different in geometry. Two of them are ground, one of them is a signal carrier. And all of this you can now simulate natively with our foundation model. And the video actually at the very beginning the loading from this geometry and structure to the moment field lines show up is actually the real pipeline in real time. So not only just the simulation piece, but the model calling, the rendering, everything else.

19:33 And you can see it’s essentially instantaneous. So, we’re really excited about this. We’re probably even more excited about this because it paints a clear path forward for this specific model architecture that I just showed you with the full 3D with the field. These are the scaling laws they hold. They’re slightly different than the ones we’re used to in an LLM world that mostly comes from a data constraint world where you sample the same data multiple times, but key messages are still for a given data size, a mod a larger model is better.

20:08 As long as there is enough data and then on the flip side, the same claim can be made as well. Given I’m short on time, I’m going to stop here. Just briefly say what’s next for us and what we’re truly excited about is obviously the scale up. That is a clear path to bigger and better. And then the second piece is the inclusion of some of those models and their derivatives that we already have into our agentic platform for true full design flows.

And then what I have not shown yet is the maronei expansion. So the inverse capability of what I just showed on the forward side but once that forward model is there that’s essentially also a reasonably chartered out path. There will be some complications I’m sure but the the path is there. So with that I just wanted to show you a little bit about how we treat or how we approach the foundation models how far we actually already are on these things.

21:07 Fully convinced with a bit more scale, with a bit more data, we’re actually at a complete inflection point when we’re talking about truly modeling physics, not the language that describes that physics ultimately. With that, thanks for the time.