Godela: Physics-informed ML model runs CFD simulations 4,500x faster than GPU-accelerated solvers

you. Hi, I'm John. Nice to meet you. What's going on? Can you introduce yourselves? Yeah, I'm Cinnamon. I'm a mechanical engineer by background. Uh built hardware for Apple, Google, Slack. Uh Wow. Hardware hardware hardware for Slack. Yeah. What do we have? RFXway. Oh, sorry. The Stanford linear accelerator center.

Oh, that not SLK. Okay, that makes more sense. I was like, what what what hardware device did I clarify? Cool. Awesome. So, I'm Abujet. Uh, I worked as a researcher at Stanford, a researcher at Harvard. I was like an intern at Intel. Um, cool. Built a lot of stuff, I guess.

Can you uh pull up on the mic a little bit more? Uh, and then tell us what your company does. Yeah, so we're Godella. We're building a frontier physics model for mechanical engineers. Okay. So currently AI models can't handle physics accurately because a lot of them are language based. Yes, ours are different.

Ours are built to handle physics accurately. Okay, which means it can be used as a faster cheaper replacement to simulations and physical prototypes. Okay, how this is something that I think a lot of labs like to to promise, right?

Solving, you know, they like to bring up Yeah, they bring up this idea of of solving these problems in future. So So walk me through. It sounds like you're actually training a model. Is it all uh reinforcement learning with verifiable rewards or are you generating a whole bunch of training data?

Is there a human data labeling component like what is the pipeline to create what you're creating? So at a core we extract embedded physics from data and that makes it generalizable. Okay. Did you want to share more? Yeah. So we don't use reinforcement learning or anything.

It's basically like this sort of like you know encoding framework like where we actually like take the mesh itself like we work with meshes right because we're like in simulation and stuff.

So we take the fluid mesh and then we like encode into like this low dimensional space and that allows to like learn the actual physics of the system. Interesting.

And this allows and when you do like symbolic regression on like the laten space you get like a lot of you know like you get a lot more generalizability than you would get with like regular models. So yeah when did you guys start working on this? Did you bring it into YC or did you pivot to this at some point?

We brought it into YC. So we were teeing a computational mechanics class together at Sanford which was teaching undergrads in mechanical engineering how to use the traditional simulation software kind of bred a hatred for those softwares.

Meanwhile Abid's doing insane research at Stanford to use ML to model the physical world in these crazy accurate ways. He's like building ML models for Intel that are replacing months of trial and error in their plasma edge process.

And I realized like hey there's this huge opportunity to out with the old out with these old simulators. Let's bring physics informed to the broader group of engineers who could really benefit from these faster, cheaper answers. Okay. CFD, computational fluid dynamics.

I have a I have an engine and I'm trying to simulate how the air will flow over the jet that I'm building. Is that an example that we could use to kind of build off of? Is it just is it just faster inference than calculating everything uh deterministically? Is that the goal? Absolutely.

Like all of simulation, every time you start with a simulation tool, the engineer started with a question, right? You've got a question in your 3D model. You want to know how is what is the drag on my Yeah, exactly. How's that going to change as I change thickness or angle of attack of my air foil?

Now, imagine instead of needing to learn a simulation software, you can ask with natural language, drop in your CAD, and get simulation quality results instantly. When we say instantly, it's 4,500 times faster than a benchmarked uh GPU accelerated solver that we tested against. Wow. Wow. Wow. Okay. Very cool.

What's the go to market like? I mean, I imagine you're selling to like very large aerospace defense companies or who else is building stuff? I mean, Apple, Google, the these companies could buy this. Yeah, exactly.

I think the huge benefit of physics informed ML is we can tackle problems that traditional simulators cannot tackle. Multi-ysics, multi scale, make it extremely feasible to tackle those problems. And we can also fill in the gaps where your idealized equations don't suffice to capture the complexity of the problem.

So these problems that Apple, Google, maybe aerospace are throwing millions at in terms of R&D and building and testing, we can give you accurate physics models that can replace your need to physically build and test a product.

So and they can probably like reality check your faster results with the traditional uh system that they have in place whenever they need to. Yeah. Build that confidence run that overnight instead of while while you're designing traction selling it. So far it's been great. So we launched two weeks ago.

So we entered a 25k year contract with an engineering firm to replace ANCIS which is $30 billion incumbent.

Um however I think our stronger pull right now is we have some exciting opportunities with enterprise customers to again tackle those highest value problems where you don't there is no solution for for modeling something like drop simulation right like there is no great simulator that gives you that fast accurate answers even though it's you know governed by software gets like 70% accuracy or something yeah interesting it's also very slow though it's like 2 weeks to compute a drop simulation on a 14-inch MacBook Pro today so like these really high value problems for enterprise customers we have the opportunity to go apply our software and they'll give give them more accurate answers.

Very cool. How big is the team? Where are you going next? How's the fund raise going? Yeah, it's three core, one advisor. Uh the fundrais is going well. We're just about Well, it's very exciting. We're just uh Yeah. Yeah. Yeah. Amazing.

Whoever whoever's like, you know, bidding, I'm sure they're, you know, going to watch this. So, congratulations on a fantastic demo day. We've been giving out hats to folks who come on the stream. Thank you so much. Great to meet you guys. Congratulations. We will talk to you soon.

Yeah, let's bring in the next team or whoever it is. And don't forget to go to vanta. com,