How Dassault Systèmes Is Building AI That Understands Physics - Ep. 296

The agents can use the virtual queen as a gym to train themselves. So the canon run, in fact, millions of simulation or design experimentation and present to you, to the human, to the engineers, to the proven solution. Welcome to the Invidia AI Podcast. I'm Noah Kravitz. My guest is Nicolas Sericeaer. Nicolas is vice-president of the 3D experience platform R&D for TSO Systems. We're here to talk about the next generation of agentic AI systems, including industry world models, virtual companions, and the systems that are driving them. Nicolas, welcome to the Invidia AI Podcast. Thank you so much for taking the time to join us. Thank you, Noah, and thank you for all of our invitation and this opportunity to be part of this podcast. Absolutely, the pleasure is ours. So maybe we can start with you telling the audience a little bit about the SO System, have a long-running partnership with Invidia. So you can speak to that a little and then also to what your role is and what the 3D experience platform is. Okay. So that's why I'm Nicolas Sericeaer, a joint SO System in 2004. And I'm now the vice-president of the 3D experience platform research and development. And you have to know that the 3D experience platform is really the foundation for our 12 brands at the SO System. You know, I think the main brands, Katia, SolidWorks, Invidia, etc. And if you don't know us, we enable our customers to imagine, design, simulate, build almost everything in the world. Cars, airplane, autonomous robots, furniture, electronic device, terapotix, med devices, etc. It's four hundred thousand customers, forty five million users, fifteen million scientists and engineers all around the world using our solution every day. And in fact, we provide our customers the factories to create their virtual twins. And what is virtual twins? It's really the scientific multidisciplinary, multiscale, they're more virtual, more real representation of the product you want to deliver. And in fact, we enable a product to be tested in the virtual world in the real condition before anything exists in the real world. And so today, my focus is on the 3D experience platform is really to transform our platform architecture into an agentic platform. And in fact, this is our shift from the SaaS platform, SaaS architecture to an agent as a service platform, to bring AI to all our customers. So what just happened in the world of AI in the past three years, in generative AI, obviously, has been, you know, this touch point that set off large language models and reasoning. And now we're talking about agentic systems. So let's talk about these two terms, virtual companions and industry world models. And what do those mean to Dyso and the Dyso world? How do you use them? And how are they different from the types of generative AI that people might be used to using for the past three years? Yeah. So let's start with industrial world model. OK, our ambition in fact is to build AI for industry. It's very, very, really important for us. It's industries. It's at the core of everything we do. And for us, AI for industry rely on three core principles. It should be grounded in science. And this is what we do for more than 40 years now. We are a science-stifted company. We deliver modeling technologies, simulation technologies. Then it should be fueled by industry knowledge. And it should be sovereign by design from the underlying infrastructure up to the model themselves. So how is it different from a generative AI? I think a classic generative AI learned the dynamics of the world from the observation and the perception of the world. So let's imagine they can see a video of a plane. They can predict if the plane will take off, if he will fly. But in fact, they don't really know why. Because they don't have the scientific explanation and the scientific foundation to understand that. And obviously, a plane does not fly by accident. So in fact, our industry world model principles, they understand how things work. They really understand the scientific foundation. They include the scientific, physical laws of the world, the physics, the engineering rules, chemistry, material science, etc. And they combine the multi-scale, multi-discipline modeling and simulation technologies we provide with AI. And the technology we are, we are devouring our industry world models rely on three technical pillars. First, the industrial knowledge. Yeah, we are talking about the standards, the regulations, the processes from the different industry we serve. And we embed the real world engineering rules. So the AI will understand and will speak the language of the industry, the jargon of the industry. Then the virtual world understanding, world industrial understanding. Here, we are devouring a ecosystem of specialized industrial AI models, which operates on our virtual twins. So the virtual and real representation of the product you deliver. Right, right. And this integrates the structure and the physics behavior. So combined with our the system, model, modeling and simulation technologies and solvers, this is how we can ensure that the AI will be grounded in science. And last is the industrial reasoning and generation. And this is where the agentech choreography takes place. And activating the industrial knowledge and the world representation to perform the experience-based reasoning. And so, about virtual companion, now, if in fact, if the industry world model provides the intelligence, the virtual companion turns that intelligence into action. What we mean with virtual companion is, we, we, we, they, virtual companion are your co-worker. Right. They understand your, your intent, of course. But they will reason with industry world models. To orchestrate, execute action in context of your business or your industry. So they will, they will comply with regulation, with your capabilities, etc. Sure. And, and, and they will protect your most precious IP, of course. And something important we don't want to replace people. We want to augment people who we want to free time to people to innovate and solve problems. So, a few months ago, we introduced a three virtual companion, porra, the business expert, Leo, the engineer, who solved complex engineering challenges, and Marie, the scientist, who bring deep, scientific expertise. So, when you're designing and deploying the virtual companions, and if we think about, sort of, a workforce, a virtual workforce of companions, that, as you said, aren't replacing human workers, but working side beside with us. In an environment like in a manufacturing environment or industrial environment where, you know, I think of my work in content, creating content, podcasting and writing. And if an LLM hallucinates, then, you know, hopefully I catch it, and I can make the correction, or maybe it inspires me to something. If a system hallucinates in an industrial environment, you know, the consequences could be much more dire. So, how do you build trust into these systems so that the people who are designing and deploying and working in these environments feel confident working alongside the virtual companions? In fact, I think the foundation for trust in our system is the scientific foundation, scientific background. Then, the human in the loop, because that's the human is accountable and remains in the loop. The choreography will pose when the human have to take decision at the critical milestone of the execution. And something very important with Deliver, and I think which is unique, is what we call IPLM, IP lifecycle management. And then, we enforce the lineage of disability, traceability, of all the interaction of AI. So, we are able to know that your content has been modified through which workflow, using what kind of models, etc, etc. And we provide, so we provide the source of trust to understand how your virtual companion be with your content. technologies in fact infuse in every layer of our architecture from Nvidia AI, with AI factories for GPUs and computing infrastructure to Nvidia AI, CUDAX libraries, and the universe technologies to accelerate AI training, inference and simulation. We focus on our partnership with Nvidia on three axes, understanding, reasoning and execution. Understanding, we integrate Nvidia NIMS models into our outscale Kubernetes platform. Outscale is our IAS to brand from the SO system. And we are huge fan on NIMS because it's super easy to deploy and perfect. Always glad to hear it. All our team are in love with the hair. So we leverage Nvidia Open Models for multi-modality, RIVA, PARS, VLM. And with PARS, we improve, for example, by a certain person, our document injection and throughput. Plus also some industry specific models such as Bayon et Mou for our virtual companion, Marie, the scientist. About reasoning now. We leverage NIMS 3 Super. And the reasoning performance for ORA, Léon and Marie, I've been improved by 20%. Without specific optimization. And this is thanks to the collaboration with Nvidia, we shared our industrial use case and benchmark. And so we were able to iterate together and to optimize the model and the integration. And then about execution. Within Nvidia, we are continuously improving the agentic execution. Leveraging the recent announcement of AIQ blueprint and deep agent. And we are also interesting and prototyping the recent announcement of NIMOClub. Of course. And we are exploring Dynamo to optimize the GPU optimization. And NIMO agent toolkits for the optimization of our agentic workflows. Can you speak a little bit to the partnership you've mentioned it as you've been talking, which kind of, you know, how it got started and more kind of what it means to diso and what it enables you to do? In fact, for over 25 years now, as you said, the system and Nvidia have redefined what is possible together. Moving from accelerating pixels to accelerating compute, computing and now to accelerating industrial AI. And so back in 20, back in 2000 from acceleration or visualization of 4th to 5th, or flagship brand and app, leveraging Nvidia GPUs to accelerating computing for similar, similar abacuse and Xflow, our simulation brand with CUDA and of course GPUs to accelerating and optimization rendering with the IRA RTX and now with the DLSS. And so this year we are opening a new chapter in this story with AI and the combining Nvidia technologies within our SredexFrance platform to deliver industrial AI platform to our customers. I want to ask you about opening proprietary models and running a hybrid model. And my understanding is that the so runs hybrid models quite a bit. Can you speak a little bit to kind of the pros and cons of each and why you go with the hybrid model so often? Say, we have an hybrid approach. Of course, we build our own models. But we want to rely on the best in class and frontier model provided by Nvidia such as the Zenimo Trond. Our optimized model by Nvidia and available through NIMS, which as I said before, enable a seamless deployment, it's super easy. Or we have also partnership with other model providers such as the Mistral. In fact, we select our models and our partners based on the performance of the model, of course, but also about the sovereignty and the regulation constraint. Okay. Because we operate worldwide, we have a customer in the whole industry and many customers are in a regulated or very sensitive industries. Sure. So we have to comply with our own regulation and all the auditability problematic, right? And so from that, we also want to calibrate the model with the customer knowledge. So we inject the industry knowledge for functioning or RIG, depending of the use case. But we, more generally, we believe in open standards. And so we embrace and we support open standards such as MCP or agent to agent. In fact, it employs our agent platform to leverage some part in the industrial system and enable, in fact, interoperable or cross system agentical geographies. I want to ask if we can dig in a little bit to a specific use case to kind of get a flavor for some of the things your customers are doing. Maybe there's an example that comes to mind you could speak to that really illustrates the use of the virtual companions and the digital platform. Yeah. I think when super cool example, I think it's a leo mechanical designer. Okay. We showcase this live, this new virtual companion in our social experience world conference last February with a gentleman attending to this conference. And so here you give leo a 3D scan or 2D drawing or a mesh of a part. It will activate the industry world model for design, orchestrate the AI model and the mod mod modding and simulation solvers. And it will perform a multi-tier planning enabling the evaluating, in fact, the mechanical interface of the part, it finds the physics, the kinematics and the design rules. And at the end, it will generate the optimized design, a physical where, manufacturable, manufacture ready, and it will do it right the first time. It's a very super example. Yeah. I think it's really illustrates our transformation from the SaaS to an agent as a service right that form. In fact, with that we are giving to our million of designers the power to innovate faster. But it's not just about speed, it's about reliability and trust. And because you know that your design works because he is born from science, from physics and he's augmented with your industry knowledge. That change that you referenced from a SaaS company to an agent, Asian as a service company. Kind of from a philosophical standpoint, I guess, or an emotional standpoint. Does it feel natural? Is it a big shift? Is it just kind of part of the way of doing things to keep innovating and delivering for your customers? And so it's just kind of the natural progression of things. How do you think about it? It's really about, in fact, with a raise of AI, we think of ourselves what is the deep impact of AI and what we do and what we deliver. What will be the new experience for the user, what will be the new technology, what we will see the cloud code, etc. What if you apply such transformation to our industrial software? In fact. It came from that, in fact, really. And so this is a lot of discussion and brainstorming at the system. In fact, we don't want to add AI on top of what we do. We want to put AI at the core. And this is why we are working with a user in the video on the different topics. What's the typical way to get started? What's the first project that a customer might typically undertake to get started with virtual companions and working with them? I think you should start from your core business and your core challenge. In fact, of course. Of course. This is where you will have attention from your teams. This is where you have your knowledge, your deep knowledge and your deep know how. And this is where this is how you know to measure the real impact of your AI and the genetic transformation. And we have an example of connecting to Leo Mechanical Design. We are working with NIR. And NIR is one of our customers working with us on virtual companion. And what they are doing to do is they recreate the virtual twin of existing aircraft. It means that they are creating thousands of parts without access to the original design. So basically, they disassemble the aircraft and recreate virtually piece by piece. So, of course, with Leo, you can imagine. Oh, we change the ZERLIFE. Yeah, automatically, the Geniathing is a 3D part from the multiple sources. That's incredible. So like everything else in technology in AI now, virtual twins, virtual companions, simulation, it's just accelerating, advancing so quickly, and obviously, agentic frameworks and models are developing just as quickly, if not faster. What's next? So what's on the horizon for the SO systems? What are the kinds of things you're thinking about? And then if you're game to take a step further, where do you think agentic systems and the idea of virtual co-workers is headed? Okay, first, I think the system, a strategy is free-aligned with the recent and media announcement about NemoClore, AIQ, all the agentic stuff. And the rise in fact of the long-running autonomous agent. And we fully agree on the associated industrial challenges, security, compliance, et cetera. And tomorrow, our virtual companion, or Alejandro Marie, we believe they will stay awake and they will continuously monitor your factory, your project execution, your supply chain in real time. And they will proactively optimize it, optimize the virtual twin without being prompt by a human. So it will create, in fact, a closed loop autonomy. And because our industry world models are grounded in physics, I think the agents can use the virtual twin as a gym to train themselves. So they can run, in fact, millions of simulation or design experimentation and present to you to the human, to the engineers, the proven solution. And you just have at the end to validate. And from that, the virtual twin, in fact, Bicom is self-evolving asset. That gets smarter day after day, in fact. - Nicolas, there's so much going on. For listeners who want to learn more, want to learn more about the 3D experience platform about deso's work with everything we've talked about, virtual companions and industry world models. Where's a good place to go? The deso website, social media, are there research papers? Where can listeners go to learn more? - Many on the deso's stem website, www.sredes.com. Oh, and I will link the page. We are communicating more and more. And yeah, thanks also to the Nvidia collaboration. We are posting more and more about what we are doing. So yeah, perfect. That's free and connect with us. - Excellent. Well, Nicolas, again, congratulations on all the work. And thank you for the years of collaboration with Nvidia. - And best of luck in everything you're doing. - Yeah, thank you to Nvidia, to the team, the incredible team.