0:15 Perfect. I’m Mike from ARENA Physica. Super excited to be here. I lead research for us. My background’s in physics originally and so I’m super excited to talk a little bit about artificial intuition and what we’ve been building over the last couple of months. I’m specifically in the electromagnetic design and engineering space. About us, ARENA Physica, our mission is pretty quite simple. We’re focusing on electromagnetic super intelligence.

0:39 And we’re already deployed with a lot of customers. So our platform actually run anything on data centers. They debug drones already. They help with eye care manufacturing on the line. So quite diverse application space, but our specialty is always electromagnetism and something electronic about a complex system that we try to understand better. We do this in three different ways. One is an agenda platform that’s Atlas at the heart.

1:08 So that helps you anywhere from design, debugging, maintenance of your complex systems. The second piece I’ll be talking a lot about today is Heaviside REM foundation model. That’s really our investment in fundamentally understanding how these systems behave from a physical point of view. And then last is actually our lab where we take the first two things and build our own phase array radar right now as well as offer services that are fundamentally transformed through the capabilities we’re building out in the other two platforms.

1:42 I’m going to start here just motivating a little bit about how we’re going and then and what we’re what we’re doing and why we’re doing it. As humans we have quite good intuition about many things and many physical systems. If you guys are looking at this coffee cup probably I’m getting slightly nervous about it tipping over. I have a sense of how it might fall. I have a sense of what happens when it falls.

2:00 I might even have a sense of like how heavy this thing is just looking at it, right? And if you’re thinking about where that comes from, that’s not something we were born with, right? But it’s probably something that you just learned through a thousands of iterations, right? Of actions you have taken and feedbacks and sensor mechanisms that you have sort of gathered, right? You got burned on the stove numerous times before you knew how that actually worked.

2:24 You jumped around as a kid on the playground to learn about physics actually feels like. So this is quite an intuitive intuitive system to us in many many ways. Jumping to electromagnetism, the world looks slightly different though. This is one simple structure that’s used in RF quite frequently. Some of you guys might know it. It’s a simple band pass filter for example. But if I asked you how does the electromagnetic wave actually propagate through this?

2:54 Probably few of you would know. Right? Even fewer people actually have an intuition of how to change the structure to have it do something different, right? There are very few people in the world actually that still know exactly how these things work and can intuitively build systems that do something with this, right? This system’s actually quite easy. It’s designed to just block some frequency and let other frequencies through.

3:18 Yet none of us could probably manipulate it just naturally to just slightly shift this frequency space. Of course I can show you what it actually looks like once we simulated it. But that’s far from building intuition, right? And even in this super simplistic case, just one layer, 8 by 8 mm, and a super simple structure, the simulation would have taken what? 4 minutes? 10 minutes? Something like this each.

3:44 Now we’re talking about one layer. New modern system like a GPU might be 20 layers, right? The size of a computer might be half a meter. So the complexity that we’re talking about here in this slightly toy example that is very real and very usable and very applicable is still orders of magnitude simpler of what some of the real world actually looks like. So even worse, why don’t we have or why don’t we have an intuition about it?

4:13 Why is EM so difficult for us to intuitively actually understand and grasp? It’s actually quite a few points that are interesting, right? If you’re thinking about the design space, it’s absolutely massive. So let’s stick with this example, almost the same structure I just showed. And if you look at it and just think about discretizing it and laying metals on let’s say a 64 by 64 grid, that space is almost infinite, right?

4:38 You’re never exhaustively look at it and actually understand it by just trial and error. Bring in connections at each and every single one of those to different layers, bring in different materials at each and every single node and you’re just as a human hopelessly lost in most cases. Doesn’t help that existing tools are unbelievably slow. We’ve already talked about this. And you guys probably all have a ton of experience setting up simulations, fiddling with meshes, all of those things, deciding what order the solver to choose and so on and so forth.

5:09 All of that stuff doesn’t help you gain any intuition either. What I think is a critical piece also, existing tools don’t learn at all. Every single time you engage with them it’s as as if you started anew, right? Doesn’t learn what you’ve done in the past or what it could have done better. It just is. It’s a tool. Nothing else. Lastly, simulations they often differ quite drastically from reality, just because you simmed one of these doesn’t mean it actually looks exactly the same once you get it back from manufacturing.

5:41 And existing I think solvers and simulators have quite a hard time integrating anything that’s not hard physics such as like contextual information in to that problem space, right? And so those I think are some things that contribute massively to us just not gaining intuition in this space at all. That’s quite a shame if you’re really thinking about it where we’re going. EM is dominating our world at the moment, right?

6:04 Pretty much every single problem and every larger trend that we’re seeing everywhere just goes to more and more EM everywhere. Starts with satellite communications, wireless, your chip interconnects and the data rate transfers that you need that are higher and higher, radar systems, the entire push is around quantum mechanics and the quantum computation. And even some of the examples of just machine perception, right? You see the rocket being caught for example.

6:34 So then how can we change that a little bit? And I think there’s two paths to more EM intuition. I think one of them we can help humans to actually get a better understanding. And I think the other one is we could build machines that just have that intuition or at least make it so that you don’t actually need it anymore because they do a large part of that job for you.

6:57 And I think at ARENA Physica we’re really trying to pursue both of those at the same time. So how do we do that? For that I’m introducing two models that we’ve been working on in our research group. That’s only one part as I said of what we do at ARENA Physica, but it’s a large part. It’s two models, Heaviside and Marconi. Heaviside is foundation model for EM.

7:19 It’s a forward model. And what I mean by that is that it takes a geometry, it takes materials, a structure, and it actually characterizes it and spits out the electromagnetic behavior that you want. So that’s very like simulator that you guys have always used, right? Every single simulator takes some input, discretizes somehow, creates some characteristic output that you would like. So that’s that’s Heaviside. The second model is Marconi.

7:45 That’s actually the fusion model for inverse design. And it does the exact opposite. It starts with EM behavior. Think about an S parameter that characterizes your system, fields, radiation patterns, or something that you actually want. And it does the opposite. It produces geometries that produce that characterization. So those two models together are actually quite powerful when you combine them. Let me say just a little bit more about Heaviside first.

8:13 It’s a foundation model. So it’s transformer based. So you guys would know it from your typical LLM sort of like architectures. It’s adjusted obviously because it’s an entirely different problem space, but it’s alike. It’s very very fast. Inference runs in something like 13 milliseconds. When we batch this up you can get this down to something like 0.3 milliseconds. So you’re you’re starting to realize about why we’re really excited about some of these systems.

8:41 It already today has sub 1 dB magnitude error. And here’s an example, same structure that I showed you earlier with the simulations. You can see Heaviside predicting the insertion losses and the and the transmission. Heaviside in red, a traditional EM solver in the gray, and then an actual test in lab as a as a white curve on top of it. So these things are fast. They’re accurate in some spaces, not not everywhere and not for everything, but they’re quite good.

9:14 This specific model has been trained on something like 2 10 million unique geometries. And in addition to that, something like 20 25 years worth of simulation data. Marconi is slightly different. To give you a little bit of an understanding of what it does, you can look at the pictures. It’s essentially a diffusion based model that many of you probably know from the image generation model. Think DALL-E, stable diffusion.

9:43 Does the exact same thing, but in this case it’s adopted to produce geometries. So you start with noise. You can condition what you want based on port placement for example or design type that you want. And then it actually just produces valid design types or design geometries that you can then actually test for their electromagnetic behavior. I want to stress this point a little bit. We’re really interested in building a foundation model, not a surrogate model.

10:14 And I think to drive home that point a little bit, I think the LLMs offer a great analogy. Before LLMs, many of you used different tools for a whole bunch of different language-based tasks that you wanted to do. Anything from translation models, separate models for sentiment analysis, and so forth. And I think even more importantly, people went through great lengths to actually introduce and analyze some of these problems first to then feed the model some more complex thing that we thought would help the model learn.

10:50 In reality, what turned out to be true is quite the opposite, right? The way we got to LLMs is by learning and looking at the more fundamental things, entire words, entire sentences, but at a much, much, much larger diversity and a much, much larger scale. And for EM, I think that will pattern will see repeat, and that’s also the approach we’re taking. We we are interested in learning the fundamental physics underneath.

11:15 So in this specific thing, EM fields is what we’re after, not the S-parameters I showed you as an example. Those are computations that follow the EM field that I learned from fundamental physical equations. How are we going to get there? I think the exact same story, diverse data and data at scale. Unfortunately for us, in EM, the world of data is slightly different and more complicated. We can’t just use the internet where we have a ton of examples, words, on how to actually do this.

11:46 But we have to build our own data factory, and we started building our own data factory. On the left-hand side, we do three things. On the left-hand side, we start with expert-guided, procedurally generated geometries. Tens of millions of these just get created, get simulated, and then actually used as training for our models. We enhance those with random, procedurally generated geometries. These don’t make any sense electronically, quite often in terms of like function, but they do teach the model quite a lot about how EM works.

12:18 And then lastly, we actually do enhance that with real measurements, where we do two things. First, make actually sure everything on the left-hand side worked out properly, the simulations run properly, that’s not always a given. And then two, give our models a chance to actually close that sim-to-real gap, right? So if we choose to, we can make a simulator not behave like a simulator, but the actual sample that you get back from a PCB manufacturer specifically.

12:51 Unsurprisingly, I would say, sort of standard sort of out-of-the-box standard models, LLMs, do really badly at this, right? I’m going to get too too much into the detail, but as you can imagine, knowing physics versus just knowing language are two separate things, right? And you see this not only in actual errors, where you see that these models perform really badly, and I have to say even comparing it is difficult.

13:15 How are you going to ask, you know, Claude to just give you an S-parameter for for for some structure, that that by itself is a difficult thing to properly compare. But you can do it, they don’t tend to do so well, not only in in in sort of error of the prediction itself, but also in cost and just in speed. But I want to flip a really bit a little bit and talk about about more.

13:42 This isn’t just research anymore. We’ve actually launched an early sandbox that you could go and try today for a limited set of these RF, radio frequency, structures that we’ve been looking at so far. And you can go and play around with it as of now. And I want to show just a few of these highlights of why I think that will lead A to a much better experience for people designing and then allow them to get more intuition as they go and do this, and B also why we’ll lead to systems that we will replace the need for you intuitively reason about it in the first place.

18:12 This is one example. The structure originally as it started on the right-hand side, the blue is the metal. P1, P2 are two ports, so we’re talking about electricity flowing between the two. And the left-hand side, you see the characteristics of that system. How does it actually behave? What do I care about is what how much gets transmitted, how much gets reflected, essentially. And what you’re seeing here is actually somebody drawing in real time, just changing the structure.

And the simulation output results that would have taken sort of minutes is actually just real-time updating on on the top left, and this is just a a screen recording. So, we used to go from having like to set up simulations and actually worrying about setting them up properly to not even having a button anymore for simulation, it’s just done automatically every single time you’re done drawing. So I think that’s quite fascinating.

Closes that feedback loop that we were talking about earlier, and I think will really lead to the fact that designers and and engineers can work with these systems and probe out and get a much more intuitive understanding of what they actually do. If we go back to our second model, Marconi, the other powerful thing is that you can do many, many of these things in parallel and iteratively.

So here you see one example, the curves on the left-hand side again are the characteristics we care about in terms of reflection and transmissions. And what we do now is we just feed our Marconi, our inverse design model, with the desire to produce that output. So give me a structure that produces that output, please. And what you can see here is a subset of the hundreds or so that it produces within a few seconds.

And then we use our forward model, our normal simulator model, to evaluate each and every single one of them against the original target we have, and you just pick the best one. So within seconds, you literally evaluated thousands, hundreds of thousands of designs, either in parallel, but the best part is not shown here. You could also just string this together sequentially to really approximate closer and closer to something like the target behavior that you actually want.

One last one more capability to highlight. We were talking a little bit earlier, we had the first talk agentic management or flows for all of this. That’s really the strength of our our platform, Atlas, that I mentioned at the beginning, is that massive contextual understanding of systems, the tools and and sort of skills that then these agents have to manage entire workflows for you. And this is actually what you can play around with on our website today yourself, if you’d like.

It’s exactly what the system is. We gave agents on our platform the tools, those two models that I showed you, to just go and play around with them themselves, right? So nothing that you see here is scripted. What you see here is a user going in and saying, “I would like to have three more designs for a bandpass filter, for example.” And everything else is agents determining themselves what that means in terms of the target I need to set, what that means in terms of the tools that I should use, and then how many of those they should kick off in parallel, for example.

So what you see on the top, you see I have a whole bunch of designs that I’ve already ran at some point from agent one, two, and three. They spit out different designs. They stopped at some point when they decided that they got close enough to the design criteria. And then the prompt is me starting another agent flow, that’s agent four, five, and six, where they just go through an additional design iteration on their own.

Now, but we can do even better, right? And so if we’re thinking about, do we even need intuition for this? That’s the that’s the last part. Now, if the system does something well enough, I might not even need to reason about it anymore or think about it too much. And this is one example, actually I have a couple of these here. You can also come and see them later on.

18:17 So these are actually real. We’ve measured them. They work. This is just one of the many examples. But if you start with a characteristic that you would like to have, you can give the system complete freedom of what it actually designs, right? So it actually comes up with things that a human couldn’t even design to begin with, right? There’s zero chance a human comes up with all of these like small little patches sort of networks.

18:41 There’s simply not enough logic, right, for a human to actually do this. But as from a system point of view, the the previous inverse design loop and evaluation loop is essentially what allows you to come up with structures like this that you can’t necessarily explain, but they have the behavior that you want. So the question then is ultimately, do I still have want to have the intuition, or am I actually just happy that the system produced what I wanted it to produce?

19:10 And all of this again is in a simple one-layer example. So even here, I would say no chance a human can actually replicate this. But if you’re now imagining again, what happens if I have five layers, two signal layers that actually couple electromagnetically, what can the system come up with that we never actually even thought of, right? What What What can it come up with that we just haven’t ever actually explored?

19:35 And that’s what we’re super excited to look at a little bit more in the next couple of weeks and months. So then, just a quick sneak preview. I would encourage you guys all to go play around with it. Have a look at what’s possible already. It’s a small sandbox, limited so that you know, we’re not getting overloaded on some things. But, where we’re going to go from here is we will do multiple layers and our goal is to tape out our own silicon this year still that is AI designed and aided in in the design process.

20:13 Generally speaking, I think there’s two big challenges that we’re going to be facing. Broader input space, the design space. If you’re thinking about what people want to design, it ranges from anything really small to antennas that are on the order of meters. So, the design space, input space into these models, and the they have what they have to handle is quite massive. So, that’s one big space of investment.

20:37 The second one that we’re going to do is on the output side. That one is also quite complicated. People are interested in anything from far-field behavior for antennas to surface currents for their PCBs. And that all of that in a on a frequency range that’s absolutely crazy, right? Anything from DC to What is it? 300 or 400 GHz at the moment for super high communication pieces. Right now, the systems that I showed you are from 1 GHz to 20 GHz.

21:05 There is quite a long range still to go to make that truly applicable and truly foundational for all applications. So, those are where we’re going. On that, I would just say thank you very much for indulging me here. Come talk to us. We’re super excited about electromagnetic behavior, building systems, building models that give us a little bit of intuition when we use them. They might be a bit better to use, easier to use.

21:32 And then also on the flip side, building systems where we don’t need our intuition at all because they do the job for us. 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.