0:00 Thanks, John, okay, good afternoon again. So I’m Matthias, I’m a CEO and co-founder of NAVASTO and an engineer, and more specifically an aerodynamicist. So we’re going to note in a second that a lot of the stuff I’m talking about is on external aerodynamics, but I like to ask you to think about the technology that’s in the back end more than about the use case I’m talking about. So I’m doing machine learning and deep learning, AI for engineering, for about a decade now, and during this time I have encountered several things again and again. And in this presentation I’d like to unpack what you mean with AI accelerated engineering, on the one inside where we’re doing it, answering three specific questions that I get over and over again, specifically: how accurate our results, how much data is needed to make good trained models, and lastly, how can we make sure that only valid predictions are being actually used for engineering decisions. And they’re going to throw in some optimization and generative AI slides in there, okay.

1:16 Before I do that, I want to start with a quick introduction of my company, for all of you out there who don’t know us. We were founded in Berlin about 10 years ago, started with two Engineers, have grown to about 20 Engineers, AI scientists, programmers, and we have developed a product, a machine learning and deep learning platform, an API solution, based on our experience of in working for the automotive industry. Solution where we decided some years ago to go for usefulness of the application more than ease of use, and this was a good decision because we ended up being used in productive processes by large OEMs, for example sead, Lamborghini, Audi Dev group, General Motors in the states are using this tool to build their vehicles. We’re very proud that we also the goto solution in Formula 1 for AI accelerated engineering, and another example is for example the america team in brania that use this tool to to design their ships. In terms of ease of use, we’ve decided to collaborate with our friends, for example at simscale, rescale and Autodesk, which make this technology then easily usable for the end users, for the thousands of Engineers that actually can benefit from such a solution.

2:51 So before I talk about the three points, like accuracy and data and stuff like that, I’d like to introduce is the idea why we’re doing it and where our vision in terms of the usefulness of artificial intelligence in engineering is. So the question is: why should we need a tool like that? At the end of the day we are having engineering processes designed to deliver a product, like a vehicle, like a heat exchanger, in a given time frame, and to arve there we have the engineering processes and the engineering capab, capability that we are having available today. So we can do this, there are nice cars out there, function heat exchanges, but the issue arises when we try to either do the same thing more quickly, meaning two examples here, in terms of two CEOs of large oems of Germany that try to accelerate the development processes of the technology on the one hand side, but even there simply with more hardware and more compute we could arrive at a point where we can actually develop those same products more quickly.

4:08 The second thing that comes in is the need for more advanced products, so to this end I’ve simply thrown in the image of the su7, aka the Porsche that was just released some weeks ago in Shanghai. So everyone needs to improve the products and and and build more advanced products, so to the end we also need more advanced engineering tools to Simply deliver better products more quickly. So this the underlying idea, and there are two ways of doing this. On the one hand side we can accelerate the workflow, like the examples we’ve seen yesterday from our friends at simscale and andera that simply improve the workflow part of the development. What we doing is more specifically going into the methods side of the development process by amending it with machine learning and and deep learning, artificial intelligence, to Simply speed up the process.

5:10 So what we do is we take geometries, like the measures, the shape of a car, and correlate those geometries with the results that we know either from simulation, so cfd simulation and FAA simulation, or from experiments. So we correlate the geometry with the known result from what has done before, train a model on that, so that we have a 3D ccate model that allows us then to test a new geometry against the Deep learning model and get a prediction of the result that you would come out of a first principle based simulation, but in real time. So you th throw in a new STL and get out the drag coefficient, the lift, the surface Fields, the volume Fields, very very quickly, and the idea here is to enable people to actually work in real time against a real time capable model. So instead of people from style, from aerodynamics, from structure, meeting in a room discussing it years, and everyone leaving the room to run a simulation, meet back a week later, we want them to be able to test an idea against the realtime capable model and make a decision based on that and iterate within those meetings. So iterate at the speed of thought is what was the idea something like 8 years ago when this picture was created in the first place.

6:35 How this looks in practice, I simply want to show you in this video. So what you see on the screen, many of you probably know blender, the open Source tool that allows to to render geometries and and modify meeses, and what we have done is we have coupled a deep learning model against the the back end of blender, so that you can now in a geometry simply modify the geometry, like like grabbing the mesh, and then instead of exporting it and running it through a first principle based simulation, like a cfd simulation, you simply query the Deep learning model in the back end about the performance of this vehicle, and you get the drag coefficient, in this case, the flow field, the surface Fields, everything that’s needed to evaluate the quality of the prediction, but in real time. So this is a real-time video, it’s literally like modif geometry, click the button, 2 seconds later you have to result here. So this is one way of using it, interactive design.

7:33 Another way of using such a technology is the combination of a predictive model together with a generative model. So the geometry I show on the left side of the slide here is created with a neural network that was trained on a set of geometries. By doing so we are including all the constraints, like the brand specific designs, like the boundaries of the design space, into the model, and then we can more or less in real time create a new geometry, which you then test against the Deep learning model, a graph neural network in this case, which gives us the performance, like the drag coefficient, the flow fields, and everything like that, so that we can then close the loop in a optimization run, like for example a circuit based optimization, or the referential evolution, to Simply search either s multiple results within a design space or find actually the aerodynamic Optimum of of a specific design space.

8:36 Another way, and this is how this technology is used most most oftenly in practice, is the conventional machine learning workflows. On the on the left hand side of the slide you see more or less the old machine learning workflow, where someone defined the design space for a specific problem, like parametric problem, then the simulations are run, then you train a model on that, and then you can test new ideas against this realtime capable model. What we see now more and more in practice is that we use not the conventional machine learning models but deep learning models, like graph neural networks or Point Cloud Transformers, to learn directly on the shape, so nonparametric shapes can be used as an input, and then we can take whatever Legacy data is available within the engineering companies and train models that then can deployed to the the end user, to the engineers, so when they have a new challenge, a new geometry, a new set of boundary conditions, they can simply use their existing workflow, the pre-processor, and then we can check if a model is available to make a real-time prediction. If that’s the case, you get your result back within seconds, and if that’s not the case, we simply switch to the existing simulation workflow, the engineer gets the result anyhow just hours later, and then we can use this exist in newly created simulation data to update the predictive model, so that becomes more and more powerful over time.

10:05 So this the use cases that you’re seeing here, and to realize those use cases what you need to have at the end of the day is three components, the three components I’ve mentioned before. So the same use case you’ve seen before, this workflow side by side, simulation and deep learning, is displayed again here just horizontally, and we have the three components: accuracy on the one hand side, second we need to make certain that we know advance in advance if a prediction can be run with high accuracy, and lastly we need to have something in place that allows us to update the models when new data becomes available.

10:41 The technology itself is physics agnostics. Al going to talk a lot about aerodynamics in a moment, it can be used and is being used for a wide range of use cases. Our customers, crashes when example, Fe is an example, internal and external aerodynamics, CHD, like the heat exchanger, Acoustics, electromagnetics. The solution at the end of the day really only correlates an input with an output, input is the shape, output is the simulation result that we know from first principle based simulation, and the the surrogated in the middle is agnostic.

11:16 Okay, if the predictions weren’t a, oh, if the predictions weren’t accurate, then all this thing would be useless. So I’d like to show you at simply a comparison backto back of a prediction and a simulation to make clear that the quality of those predictions is really amazing. To the end I based this part of the presentation of a data set of the driver, it’s a European research vehicle, you see the Baseline on the left hand side of the slide here, in the geometric variation that went into this data set on the right hand side, so those are the extremes of the geometry that were put in there. And we had a lot of those simulations, so in total 500 simulations, we took a subset of the data to train a model, and the remaining something like 150 simulations to validate the model against unseen geometries, and the result of that is what you see on the slide here.

12:18 So we compare the drag coefficient of this vehicle against the the ground Ruth from cfd simulation, so the x-axis is the ground Ruth, this is what comes out of an 8h hour simulation, the y-axis is our prediction, the real time result for new geometries, and everything that’s on this diagonal is a perfect prediction, meaning that the Deep learning model exactly reproduces what you would get out of a cfd simulation. And as you see here, we are about in the order of magnitude of an error that would come from a first principle based simulation, so we are in a noise of what you see in such a simulation.

13:02 This is of course not limited to scal of values, but we can also predict surface fields and volume fields. In this case, on the left hand side you see the result from the first principle base simulation, on the right hand side our prediction, pressure on the surface, velocity behind the vehicles. This is the way that the Formula 1 teams like to look at this Pro problem, and the interesting thing about this slide or this animation is that the update rate of this video, 70 M seconds per frame, is exactly the inference time that we have in this case for the Deep learning model, while each of the simulations shown on the left hand side of the slide take about 8 hours to come back on a GPU solver.

13:45 So zooming in, simply comparing in more detail the quality of the solution, again on the left hand side the truth, on the right hand side the prediction, you see that it’s very very close to each other. Even if you look into further details, like I now plot the velocity components of this field, on on the screen you see that both the location and the magnitude of the solution matches very well, and even if they put the two of them on top of each other, you find that the prediction is more or less in distinguish from what you would get out of a simulation, just that it’s real- time capable.

14:28 To prove the point that this is not limited to external aerodynamics, I brought two more use cases. This one here is a crash use case, this is the crash box of a vehicle, on the left hand side the simulation running for at least a few hours, on the right hand side the prediction, which is happening in about 20 milliseconds per time step. The same thing here on a full vehicle, on the left hand side, in this case our prediction with a about 20 milliseconds per frame, on the right hand side several hours for the for the the full physics based simulation.

15:01 And finally one example, the only thing that I found within my company with relation to additive manufacturer, it’s a lest structure or rather hyper structure, we have done this work together with aimus who produce the data. You see the simulation result from FEA on the the middle column, the prediction that we make, and I think what you see on the right hand side it is video, I mean this is not even measurable. The prediction time of such a simple Network is one of the options that we see is coming out of those realtime capable models. Even if you don’t change the geometry, you could test a lot of load cases very quickly to ensure, for example, the robustness of a specific design.

15:55 Okay, now this part was at the end of the day about is it possible to train machine learning models and deep learning models that accurate. The answer is is yes, we can train models with some Asis attached here that are producing results indistinguishable from a simulation. The question is now how much data is needed, and I’ve shown it before a model that was trained on 350 engineering simulations, usually starting point in our business is something like 50 to 100 to train a model, but at the end of the day it’s really about take whatever is there to train a model and then continue to teach the model over time when new data becomes available.

16:36 And I’d like to Simply illustrate this based on an example of the driver you’ve seen before, and then I going to add the information of the cheda to the same model. So as Illustrated before, we can train a model, we just take whatever stair are able to predict the drag coefficient or able to predict the flow field and the surface Fields. Okay, now a new data set becomes available, in this case in blue you see the driver model, in Orange it’s the VW cheda, so it’s a production vehicle in contrast to the research vehicle. This has an engine bay, this is an exhaust system, both are limousines, both are cars, but they are distinctly different in terms of geometry. Now however, if you would run an inference on the model that was trained on the driver with the cheda, the output would be nonsense, a high nonsense. However, if you now add a few number of samples, like 20 in this case, to the pre-trained model, we were able to predict new variations of the cheddar, this are the blue dots in the bottom left corner of the plot, and still have a model that’s able to predict to predict the shapes of the driver and everything in between. So the underlying idea is here that you have a model that simply grow over time, whenever new data becomes available you can add this data to the model and have a growing model within the coll, the corporation, that’s fed by every simulation engineer that is using a simulation workflow.

18:09 So the two points were now accuracy and amount of data required, my last point I’d like to make is how can we make sure that only valid predictions are used today base engineering decisions upon. There is no analytical way to quantify the error of a machine learning or deep learning prediction, we need to make some metrics available nevertheless for the engineers to judge if they can base an actual decision on this prediction, and to this end we are proposing to do two things.

18:42 The first thing we usually do is we compare the new challenge, the new problem, the new geometry, to everything that has been used in the training before. So on the left hand side you see again my driver vehicle, now with a morphing roof height, everything that looks like a normal car was part of the training data set, everything that looks like a flat thing, or I don’t know, like a van, was not part of the training data. And what we then do is we simply compare the new shape and the boundary condition with the data set that was available during the training, and we can then tell the engineers if they’re running interpolation or extrapolation. Interpolation works like a a charm, like you’ve seen before, extrapolation is definitely luck involved, so usually we have something like a confidence value of 7.8, where Engineers are able to use the prediction, for everything else they’re required to run a simulation. So it’s the one thing we can do.

19:43 Another thing we can do is we can use the statistical properties of the machine learning and deep learning model to also compute local uncertainty within a model. So on the left hand side of my car you see color encoded the uncertainty that’s coming out of the machine learning model, different variant available, multi-head networks, deep unsolvable aaging drawboard, whatever is available in the world of machine learning deep learning, can be used to not only get a average value but also some form of standard deviation or distribution of data to estimate the error. So red means the model tells me it’s rather uncertain, blue means the model tells me I’m certain, so this is an information I get without running a simulation. But if we would then run a ground truce simulation, meaning I simulate the thing and subtract my prediction from the actual value, I get an error on the surface here, red again meaning High error, blue meaning low error, whatever the specific error is, you see that my initial uncertainty estimation correlates very well with the actual error I encounter for the gr through simulation.

21:00 So that I have a set of tools to actually judge if I can trust this prediction or not, and thus it’s not just guessing if it’s useful or not, but there are several quantifiable measures available to make sure that only Val predictions are used to based engineering decisions of. Well, with that, I actually, I want to close the presentation, I hope it was somewhat useful. If there are any questions I’m more than happy to answer them.

21:31 How input do you have on your model? That, absolutely, that depends on the GPU that you have available, so in practice we are working with something like 1 million or two million notes on the geometry, million, million, think, think about a Formula 1 car. Formula one car, you need to represent the entire geometry with all those wings and Vortex generators, so every coordinate point isut. Yes. Does the mes need before and after when you TR? No, this method, they’re different methods, like a GNN, like a point CL Transformer, those are invariant, no the the mesh can be arbitrary in the input. Does it make sense? That’s too, no, that’s that’s not too big, this is in production, this is not research, we are we are selling this. Thank, it’s really good.