Right. All right. Awesome. Well, thanks very much for the intro. My name is Pablo. I obviously work for Nvidia. I’m a senior solutions architect specializing on CAE and AI physics. And my role here at Nvidia is really to help developers in the space on CAE and AI physics integrate some of our libraries and I’ll talk about those today. As well as generating training material and one of the things that I really focus as well is bringing all the feedback from our developers and our partners back to the product teams developing those libraries.

0:52 So obviously as you probably all know the engineering space is you know up for a revolution. The pressure on engineering companies to accelerate these design and manufacturing workflows is greater than ever and the need to have end-to-end digital twins is imperative and today I’m going to be talking about how we’re seeing the use of AI physics and AI agents accelerate the the space. So maybe before before starting on that I just wanted to give you a bit of an overview of how our team at Nvidia and industrial engineering see our contribution in the space.

1:31 So first of all you know we see it in three kind of pillars. The first one that you probably know most is the core workload acceleration. So we work a lot with some of the software vendors and see Siemens and all those to accelerate principal solvers. So use GPUs to accelerate the the the solvers. The second pillar that we focus a lot of is engineering with AI and this is the one I’ll focus today.

1:58 So all about using or developing AI surrogates and use of agents to accelerate end-to-end engineering workflows. And then the latter one is real-time digital twins. So the idea of getting all the CAD design all the engineering data solver data simulation data and bringing it into a single 3D interactive platform and this is where our Omniverse libraries come in. For many of you you won’t actually make use of those libraries because those libraries we normally with you know the the software vendors so that you make use of them such as with Siemens as I as I said and many others.

2:40 If you’re interested in the core workload acceleration and you work on let’s say developing new new solvers and you’re interested in learning more about our CUDA-X libraries and CUDA math libraries I really do encourage you to give a listen to my colleague Ian’s presentation in last year’s CDFAM in New York because he goes into much more depth on things like CUDA sense and AMGX so all our accelerated matrix solvers.

3:04 But as I say today I’m going to be focusing more on the the engineering with AI. So My I used to work as an engineer. I used to work as an engineer in Formula 1. So this this is kind of like my way of putting together the the engineering cycle. So as an engineer you would normally be doing things like the CAD design the pre-processing simulating and ultimately you know post-processing and so analyzing the results of the simulation right and we see that this is the way of doing traditional engineering and you would normally loop through this design cycle quite a few times.

3:41 Now we’ve seen and it’s really now a fairly mature industry the use of AI in some of these steps right so we’ve seen topology optimization and generative design. There’s a lot of also let’s say really interesting research and applications around the simulation space so the use of AI surrogates to basically predict the output of your numerical simulation. AI preconditioning as well is something that’s becoming quite big so the idea of using a surrogate to initialize your principal solver so that converges quicker.

4:15 And then obviously a lot of work as well on inverse design. So if you manage to have a a surrogate that is differentiable which is the case normally for your most machine learning models then you can basically optimize with respect to your parametric design right? Now here at Nvidia we’ve done a lot of work on the simulation space right and that’s why we developed our physics Nemo platform.

4:40 So for those of you who may not be familiar with it physics Nemo is a platform which is a collection of open source GitHub libraries really where you have data pipelines to do basically all the extract transform transform and load of your simulation data so that you get it into an AI ready format optimized architectures so we we basically get most of the state of the art architectures for physics AI and optimize them and bring them into a single library.

5:12 Physics AI integration so if you’re interested in doing like PINs and so on we do have let’s say ways to encode those physical laws and then as you can imagine distributed scaling and that’s all that Nvidia let’s say we focus a lot on how can we scale this across multi-GPU and multi-node. And we’ve done a lot of work on this not only on external aerodynamics but other fields such as crash subsurface simulations additive manufacturing you name it.

5:42 It’s surprising that the architectures are fairly let’s say similar across those. So I’m I was told that the audience was fairly technical so I thought I’d take a bit of time to walk you through how a physics AI workflow works in practice and maybe just dive a little bit deeper into into that right? So I’ll give you one example one real example. So suppose I’m interested in working on crash crash simulations and developing surrogates for that right?

6:16 So for we start with the simulation data set. So in this case for instance I’m working with a bumper beam crash data set that I generated using the radius package. And I generated 135 variations with different basically geometries initial and boundary conditions. So basically the speed at which the bumper crashes against the rod. The next thing is really to take the simulation data and prepare it to basically be ingested in physics Nemo so basically under the hood PyTorch right?

6:52 So that’s why we developed the physics Nemo curator library. So it’s a it’s a again open source library where you have domain specific readers and transforms for many domains. So if you’re working with aero data and you’re you’re basically loading STLs VTUs VTPs and so on so we have readers for those. If you’re working with crash or basically finite elements D3 plot rad files K files you name it and a few others around different different industries right?

And it’s really a convenience library. You have GPU accelerated data processing for things like constructing graphs and going over graphs that those kind of things suit quite well to be accelerated for the GPU. And the idea is that since we work as well on the training we know effectively what the contract looks like and how can we optimize the data formats to basically accelerate the training as well.

7:49 So that’s why you know normally one of the things that we do is that we convert basically your raw simulation files into something called for instance like a Zarr file. So Zarr is nothing more than an open source format cloud-native and basically allows you to chunk your simulation data into time and space which is quite convenient later on because it allows you to do basically what we call lazy loading.

8:11 So the PyTorch data loader then just basically during training fetches what it needs and this is actually quite convenient and really important because as I mentioned here is just maybe a small example right with 10 gigabytes per case but these methods we’ve done it with things like aero volumes which might be 200 million cells plus and 50 gigabytes per file. So optimizing the data formats and the data loaders is quite a big thing.

8:39 The next thing is actually going into the training right? Physics Nemo is our library for the training side of the of the world and we’ve got as I mentioned a whole bunch of model architectures and data ingestion pipelines optimized and what people maybe don’t realize is actually we invest quite heavily into this and that’s why we also have a whole team of research scientists developing the new the newer generation of physics AI models.

9:04 And one such example is actually the year transform over. It’s a paper we released I think it was December last year and it just shows you well it’s a transformer based architecture. I won’t go into too much more of the detail but the but it’s a really interesting way of adapting transformers for physics AI and for some of the domains it’s it’s close if not state of the art.

9:30 So really interesting research as well. So once you’ve trained the model obviously I’m going very quick through the workflow but then you can make inference on the model. The inference will normally take you know for this case under a second and you can see that even just with 135 cases a data-driven model like this one can get fairly good correlation in displacements and things like plastic strain and so on.

9:53 So I guess maybe a caveat here, right? Everyone asked me after this slide is how good are these models, right? When you don’t encode physical laws and so on. The answer really depends on on your data, right? But and how much data, how varied the data is. But we’ve seen that the newer generation of models are getting increased geometric awareness, physical consistency, and for me probably one of the most important thing is that they’re starting to scale really big.

10:25 So, it’s kind of scaling towards industrial scale simulations where you have millions and millions of of cells. Having said that, it’s worth saying that Physics Nemo is still a developer library, right? So, we focus on building Physics Nemo for developers and making it as modular as possible so that you can pick and choose and bring and use what you want, right? So, we have partners who maybe only using the Physics Nemo Curator library for the data processing, and they develop they develop their own models.

10:56 Others might be using perhaps just even the encoder of a model, right? It really depends. So, we work very closely with some of our partners here as well. So, we’re really proud to to to work with them to integrate Physics Nemo into their platform. So, for instance, some work that we did with SimScale where we were working on centrifugal pump optimization. We’ve seen as well with Luminary Cloud, with Cadence, so many more.

11:21 But yeah, my point here is that we really do focus on on developers here, and we want the developers to take it, dissect it, and extend it as they wish. So, up to now I really talked about basically using AI really in the middle, but when we talk to some of our end customers, right? So, some of the engineering companies, they they do tell us as well like it’s not enough, right?

11:49 So, even if you have AI in the simulation, the engineer still faces a lot of bottlenecks, right? So, the engineer is normally using 10 different software packages, bringing data across and the the process is still limited by an engineer driving the process from beginning to end, right? So, that’s why a lot of our recent focus has been on the use of agents. So, they use agentic AI to do things for instance like intelligent CAD preparations.

12:19 So, can an an agent go in, use your CAD software, fix the geometries, intelligently case sorry, setting up your case given that it knows the documentation and the requirements, and for instance doing thing as well as automated reporting. So, let me go into a bit more depth. This slide maybe a bit you probably know what an agent is, really. But at its core is nothing more than a system with a reasoning LLM, access to tools, so it can basically call APIs, MCPs.

12:52 It can make use of your computer, and it also has access to files and memory, actually. Probably the most important bit of it where it has access to let’s say long-term memory that you provide via rag, but also session memory and user memory, and it can evolve over time, right? To give you a feel of what we work at NVIDIA, so we try to provide with basically the whole stack from beginning to end when it comes to developing agents.

13:18 So, not only the AI infrastructure that you probably know quite well, but also a lot of the work that we do is on the NIMs. So, NIMs are optimized inference microservices. Think of them as basically containers. So, pre-packaged containers with very much optimized inference. So, for instance, there is NIMs for pretty much every open state-of-the-art LLM, right? And we take those LLMs, we package them into a container, we have a software team dedicated to optimizing every layer so that regardless of the hardware you’re running it, you’re maximizing the the inference efficiency.

13:54 Then we get the Nemo libraries. So, those are basically used to adapt and fine-tune some of our some of the agents and some of the LLMs. So, the curator, curator, evaluator, retriever, and so on. And then the AI blueprints on top. So, most basically close to the application layer where you have basically reference architectures for building multi-agent systems. So, how can you connect the dots? How can you introduce tools?

14:15 How can you introduce skills? And to do some tasks like research assistant, and One thing I did mention here, that’s why I put it there, is that the the research assistant one is a really cool one. Actually, it’s called the AI Q research deep research assistant, and it’s actually ranked number one in the open deep research benchmark. So, one thing I would say is even if you’re not interested in deep deep research with agents, it’s a really really cool example of multi-agent systems.

14:52 It’s actually developed by some of our Kaggle grandmasters. So, yeah, really good insight into how to build a multi-agent systems there. One thing that we do as well is we train our own set of large language models, and this is the the NemoTron initiative. So, for instance, NemoTron 3 is our latest generation of family of models trained specifically for agentic tasks. For So, for instance, NemoTron 3 Super, which was released just before GTC, ranked number four just behind some of the closed frontier models in agentic basically open source agentic benchmark for in this case I think it’s OpenFlow, right?

And we’re actually training post-training NemoTron Ultra as well. So, we really do hope that that’s going to be, you know, very much up there. And the reason we do this is very much that we want as NVIDIA to learn and be very much close to what where the LLM industry or where the technology is heading, right? So, then we can inform better our product team and our hardware and basically design the most efficient hardware.

16:04 And the reason I think I also bring it up is I guess most of you are doing basically working with engineering companies or at an engineering company, and from my experience when I talk when I talk to these companies is data is their IP, right? So, their design is their IP, and there are certain conditions where the data cannot be sent to a cloud provider, right? So, some engineering companies will have to host and basically run their LLMs on premise.

16:31 So, the idea of having a an open source LLM that you can run on premise, I think and it’s close to the state-of-the-art, I think is really interesting. And it’s as I say, it’s fully open source. We open source the data, we open source the the training codes, the inference code, the post-training, you name it, right? So, it’s really easy to adapt to basically and fine-tune to your specific use case.

16:57 Now, let me take some time to show you some examples. I think interesting examples. So, we’ve done a lot of work on agents internally as well on the CAE workflows. So, from basically working with reservoir simulations, driving post-processing with Omniverse, even developing coding agents for air physics, but let me just, you know, be a bit more conscious on time. So, for instance, this is actually an example that we did with a company that does reservoir simulations.

17:24 So, in the oil and gas industry. And the video doesn’t do it justice, but the idea, and this is going to be open source. The idea is that you have your case data, the engineer can bring in the the case data, ask the agent to run a simulation, and then the simulation will actually sorry, the agent will actually go inspect the case files, understand that there’s an error for instance in the well dimensions, run the simulation.

17:48 There’s a agent looking at the logs, making sure that the simulation converges, and then even doing the post-processing. So, again, it’s very much simplified in the video, but the idea is that you have multiple agents coordinating things from preparing the case, analyzing the basically the the logs of the simulation, checking for conversions, and so on. So, this is going to be open source fairly you know, in the next in the next month I think I think.

18:17 We have other other examples as well in the industry. So, for instance, we’ve worked with NASCAR doing agentic post-processing, with Dassault Systèmes, with Rescale and McLaren. Let me show you this video. I think there is music to the video. Well, that’s fine. So, this is actually an example with NASCAR where they do a lot of their post-processing. So, with Omniverse and Kit CAE. And the idea is that they we know we’ve now helped them do this basically proof of concept where you have an agent that can drive the post-processing.

18:50 So, creating like new slices. There’s actually even a VLM, so a visual language model that can capture the viewport and do things like, you know, telling you where the the vortex structures are, where the flow separation is, and even generating for you reports as well. So, this is actually again, the beginning of I think quite an exciting collaboration an exa- an really an example of how we believe that agents can can help for instance in the post-processing.

19:21 This is another one that we did with Dassault Systèmes in CATIA. I’m sure a lot of you work with CATIA, and I’m not sure if in the desktop or the cloud version. But for instance, with Dassault Systèmes, we’ve worked in integrating our NemoTron models in an agentic system so that for instance in this case, the user is asking the agent to reconstruct this part in the in the landing landing gear system, and basically use basically the native CAD features.

19:47 So, it’s under the hood is basically calling all the APIs from Katia, reconstructing the geometry, you know, things this would take an engineer at least an hour, right? And then for instance running all the the stress simulations, making modifications as you can see here as well. And this is all driven by the agent, which I think is really exciting. And then running or re-running the stress simulations again, and even generating engineering designs.

20:16 So this is actually a part of a quite an ongoing collaboration with the sole systems and I think an example of how this would take an engineer a couple of hours and with an agent we can do it in in a matter of minutes. Obviously I couldn’t leave today without at least mentioning Open Claw. I think Open Claw has really shown us how we can basically develop these long-running agents with autonomous persistence that can basically update themselves and this really has unlocked a lot of productivity but it it comes as well this degree of unbound autonomy comes again with some security implications.

20:56 So that’s why you know, we developed Open Shell. So Open Shell actually it’s it’s very new. It’s it started in we announced it in GTC and the idea is that Open Shell sits between your coding agent so that could be basically Open Claw or could be Cloud Code or whatever you’re using and the application layer so your file system and so on and basically controls what the agent can see, what it can do and what it can call, right?

21:23 And it does that with deny by default and it it’s really important if you’re developing these agents in production. I may be behind on time so I’ll be very quick on this one. I know one of the things that we also implemented with the Nemo Claw so basically Nemo Claw is a reference implementation of Open Shell with Open Claw and it includes all the let’s say the guardrails, a privacy router as well.

21:51 I think that’s really interesting if you’re working in CAE because one of the things that we implement and these become quite sophisticated is that we can add basically conditions and basically intelligent routing. So if you’re working with sensitive data you route the data to like a local model and if you’re basically doing something that is much less sensitive you can route it to one of the cloud cloud frontier models.

22:12 So this is actually something that we’re seeing quite a lot. So I will skip the slide. Yeah, the idea is that once we have all these I think the the opportunity is is tremendous, right? So I think we really want to work with developers and partners in the space to think beyond just applying agents to a single step, right? And looking into multi-agent architectures and we do see a future where as an engineer you could basically ask a claw to basically take your drone, optimize the blade and then the claw would basically design a plan, call a design claw that would interact with your CAD software, call a simulation claw that will run your simulation and check for convergence, an analysis claw that would basically analyze the whole thing and decide on the next step.

23:03 So as I say this is a look into the future. It’s still not there but I’m sure we can you know, work together on this and that’s why I like to finish the slide or my presentation with this slide here which is just a big thank you to all of our ISVs and developers that we work very closely with them. And yeah, it’s really exciting times and I’m sure we can use our technology to basically help the whole engineering ecosystem above.

23:30 So with that I thank you very much. This is my email as well if you need anything and Ian’s email so this is Ian works on developer relationships. So anything you need you feel free to to reach out to us because yeah, we’re more than happy. So thank you so much for your time. To learn more about the CDFA Computational Design Symposium, access the archive of previous presentations, interviews with speakers and information about future events around the world, visit CDFAM.com.