How Dassault Systèmes Is Building AI That Understands Physics | NVIDIA AI Podcast Ep. 296

0:00 The agents can use the Virtual Twin as a gym to train themselves.

0:05 So they can run, in fact,

0:07 millions of simulations or design experimentations and present to you,

0:12 to the human, to the engineer, there is a proven solution.

0:19 Welcome to the NVIDIA AI Podcast.

0:21 I'm Noah Kravitz.

0:22 My guest is Nicolas Cerisier.

0:24 Nicolas is vice president of the 3DEXPERIENCE

0:27 platform R&D for Dassault Systèmes.

0:30 We're here to talk about the next generation of agentic AI systems,

0:33 including industry world models, Virtual Companions,

0:37 and the systems that are driving them.

0:39 Nicolas, welcome to the NVIDIA AI Podcast.

0:42 Thank you so much for taking the time to join us.

0:44 Thank you, Noah, and thank you for the invitation

0:46 and this opportunity to be part of this podcast.

0:49 Absolutely, the pleasure is ours.

0:51 So maybe we can start with you telling the audience a little bit about Dassault

0:56 Systèmes—have a long-running partnership with NVIDIA,

0:58 so you can speak to that a little,

1:00 and then also to what your role is and what the 3DEXPERIENCE platform is.

1:04 Okay, so I'm Nicolas Cerisier.

1:07 I joined Dassault Systèmes in 2004, and I'm now the vice president

1:12 of the 3DEXPERIENCE platform research and development.

1:14 And you have to know that the 3DEXPERIENCE platform is

1:18 really the foundation for our 12 brands at Dassault Systèmes.

1:22 You know, I think the main brands CATIA, SOLIDWORKS, SIMULIA, etc.

1:28 And if you don't know us, we enable our customers to imagine,

1:34 design, simulate, build almost everything in the world.

1:38 Cars, airplanes, autonomous robots, furniture,

1:43 electronic devices, therapeutics, med devices, etc.

1:48 It's four hundred thousand customers, 45 million users,

1:54 15 million scientists and engineers all around

1:58 the world using our solution every day.

2:01 And in fact, we provide our customers

2:03 the factories to create their Virtual Twins.

2:07 And what is Virtual Twins?

2:09 It's really the scientific, multidisciplinary, multiscale, V plus R,

2:16 virtual plus real representation of the product you want to deliver.

2:22 And in fact, we enable a product to be tested in the virtual world,

2:26 in the real condition before anything exists in the real world.

2:33 And so today, my focus leading the 3DEXPERIENCE platform is

2:36 really to transform our platform architecture into an agentic platform.

2:41 And in fact, this is our shift from a SaaS platform, a SaaS architecture,

2:45 to an agent-as-a-service platform to bring AI to all our customers.

2:52 So much has happened in the world of AI in the past few years,

2:55 and generative AI, obviously,

2:57 has been this touchpoint that set off large language models and reasoning,

3:01 and now we're talking about agentic systems.

3:05 So let's talk about these two terms,

3:07 Virtual Companions and industry world models.

3:10 And what do those mean to Dassault and the Dassault world?

3:14 How do you use them, and how are they different from the types of generative AI

3:18 that people might be used to using for the past few years?

3:22 Yeah, so let's start with industrial world model.

3:27 Our ambition, in fact, is to build AI for industry.

3:31 It's very, very, really important for us.

3:34 Industry is at the core of everything we do.

3:38 And for us, AI for industry relies on three core principles.

3:43 It should be grounded in science.

3:46 And this is what we do for more than 40 years now.

3:48 We are a scientific company.

3:51 We deliver modeling technologies, simulation technologies.

3:56 Then it should be fueled by industry knowledge.

4:00 And it should be sovereign by design—from

4:03 the underlying infrastructure up to the models themselves.

4:08 So how is it different from a generative AI?

4:10 I think a classic generative AI learns the dynamics

4:14 of the world from the observation and the perception of the world.

4:20 So let's imagine they can see a video of a plane.

4:26 They can predict if the plane will take off,

4:30 if it will fly, but, in fact, they don't really know why.

4:35 Because they don't have the scientific explanation

4:37 and the scientific foundation to understand that.

4:41 And obviously, a plane does not fly by accident.

4:47 So, in fact, our industry world model principles,

4:51 they understand how things work.

4:56 They really understand the scientific foundation.

4:59 They include the scientific, physical laws of the world.

5:03 The physics, the engineering rules, chemistry, material science, etc.

5:09 And they combine the multi-scale,

5:11 multi-discipline modeling and simulation technologies we provide with AI.

5:17 And the technology we are delivering,

5:19 our industry world models, rely on three technical pillars.

5:25 First, industrial knowledge.

5:27 Here we are talking about the standards, the regulations,

5:29 the processes from the different industries we serve.

5:33 And we embed the real-world engineering rules,

5:36 so the AI will understand and will speak the language of the industry,

5:40 the jargon of the industry.

5:42 Right, right.

5:43 You see?

5:44 Then the virtual, the world understanding, the world industrial understanding.

5:50 Here, we are delivering an ecosystem of specialized industrial AI models,

5:55 which operate on our Virtual Twins.

5:59 So, the virtual and real representation of the product you deliver.

6:03 Right, right.

6:04 And this integrates the structure and the physics behavior.

6:09 So, combined with our Dassault Systèmes

6:11 modeling and simulation technologies and solvers,

6:13 this is how we can ensure that the AI will be grounded in science.

6:20 And last is the industrial reasoning and generation.

6:24 And this is where the agentic choreography takes place,

6:28 and activating the industrial knowledge and the world

6:32 representation to perform the experience-based reasoning.

6:37 And so, how about Virtual Companion now?

6:40 If, in fact, if the industry world model provides the intelligence,

6:43 the Virtual Companion turns that intelligence into action.

6:49 What we mean with Virtual Companion...

6:52 Virtual Companion are your coworkers.

6:56 They understand your intent, of course,

6:58 but they will reason with industry world models to orchestrate,

7:03 execute action in context of your business, of your industry.

7:07 So they will comply with the regulation, with your KPIs, etc.

7:14 And they will protect your most precious IP, of course.

7:18 And something important—we don't want to replace people.

7:21 We want to augment people.

7:23 We want free time for people to innovate and solve problems.

7:28 So a few months ago, we introduced three Virtual Companions.

7:32 AURA, the business expert.

7:33 LEO, the engineer who solves complex engineering challenges.

7:37 And MARIE, the scientist who brings deep scientific expertise.

7:41 So when you're designing and deploying the Virtual Companions,

7:47 and if we think about sort of a workforce,

7:50 a virtual workforce of companions that, as you said,

7:53 aren't replacing human workers, but working side-by-side with us,

7:57 in an environment like in a manufacturing

8:01 environment or industrial environment where...

8:03 You know, I think of my work in content,

8:06 creating content, podcasting and writing, and if an LLM hallucinates,

8:10 then, you know, hopefully I catch it,

8:12 and I can make the correction, or maybe it inspires me to something.

8:16 If a system hallucinates in an industrial environment,

8:19 the consequences could be much more dire.

8:21 So how do you build trust into these systems so that the people who are

8:28 designing and deploying and working in these environments

8:30 feel confident working alongside the Virtual Companions?

8:35 In fact, I think the foundation for trust

8:38 in our system is the scientific foundation,

8:43 scientific background, then the human in the loop, because at the end,

8:49 human is accountable and remain in the loop,

8:51 and the choreography will pause when humans have

8:54 to take decision at the critical milestone of the execution.

9:01 And something very important we deliver,

9:03 and I think which is unique, is what we call IPLM,

9:08 IP Lifecycle Management, where we enforce the lineage,

9:13 auditability, traceability of all the interactions of AI.

9:18 So we are able to know

9:21 that your content has been modified through which workflow,

9:25 using what kind of models, etc.

9:29 And we provide the source of trust to understand

9:36 how your Virtual Companion behaves with your content.

9:40 So NVIDIA is bringing technologies,

9:44 open models, Omniverse, accelerated computing,

9:47 AI physics libraries, all these technologies into the stack.

9:51 How do technologies like these help enable

9:54 more capable and more secure agentic workflows?

9:59 So NVIDIA technologies, in fact,

9:59 infuse in every layer So NVIDIA technologies, in fact,

10:01 infuse in every layer of our architecture—from NVIDIA AI

10:05 with AI factories for GPUs and computing infrastructure to NVIDIA AI,

10:12 CUDA X libraries, Omniverse technologies to accelerate AI training,

10:19 inference, and simulation.

10:21 Regarding NVIDIA AI and agentic,

10:25 we focus on our partnership with NVIDIA on three axes.

10:30 Understanding, reasoning, and execution.

10:33 Understanding—we integrate NVIDIA NIMs models

10:36 into our OUTSCALE Kubernetes platform.

10:40 OUTSCALE is our IaaS.

10:42 It's a brand from Dassault Systèmes.

10:44 And we are a huge fan of NIMs because it's super easy to deploy and...

10:50 perfect.

10:50 Always glad to hear it.

10:52 All our team are in love with NIMs.

10:54 Awesome, love to hear it.

10:56 So we leverage NVIDIA open models for multimodality— Riva, Parse VLM.

11:02 And with Parse, we improve, for example,

11:05 by 30%, our document injection and throughput.

11:10 Plus also some industry-specific models,

11:13 such as BioNeMo for our virtual companion, MARIE, the scientist.

11:18 About reasoning now,

11:21 we leverage Nemotron 3 Super and the reasoning performance for AURA,

11:28 LEO, and MARIE have been improved by 20% without specific optimization.

11:39 And this is thanks to the collaboration with NVIDIA.

11:42 We shared our industrial use case and benchmark.

11:46 And so we were able to iterate together

11:48 and to optimize the model and the integration.

11:51 And then about execution.

11:54 With NVIDIA, we are continuously improving the agentic execution.

11:59 Leveraging the recent announcement of AI-Q Blueprint and Deep Agent.

12:05 And we are also interested in prototyping

12:10 the recent announcement of NemoClaw, of course.

12:14 And we are exploring Dynamo to optimize the GPU optimization

12:19 and NeMo Agent Toolkit for the optimization of our agentic workflows.

12:25 Can you speak a little bit to the partnership?

12:28 You've mentioned it as you've been talking, but just kind of...

12:29 You've mentioned it as you've been talking, but just kind of...

12:30 how it got started and more kind of what it

12:34 means to Dassault and what it enables you to do.

12:37 In fact, for over 25 years now, as you said,

12:41 Dassault Systèmes and NVIDIA have redefined what is possible together.

12:46 Moving from accelerating pixels, to accelerating computing,

12:51 and now to accelerating industrial AI.

12:55 And so back in 20...

12:58 back in 2000, from acceleration of visualization of CATIA V5,

13:04 our flagship brand and app, leveraging NVIDIA GPUs,

13:09 to accelerating computing for SIMULIA, Abaqus, and XFlow,

13:15 our simulation brand, with CUDA and of course, GPUs.

13:21 To accelerating and optimization rendering with Iray, RTX, and now with DLSS.

13:28 And so this year, we are opening a new chapter in this story with AI

13:35 and combining NVIDIA technologies within our 3DEXPERIENCE platform

13:40 to deliver industrial AI platform to our customers.

13:45 I wanna ask you about open and proprietary models and running a hybrid model.

13:51 And my understanding is that Dassault runs hybrid models quite a bit.

13:55 Can you speak a little bit to kind of the pros and cons

13:59 of each and why you go with a hybrid model so often?

14:05 Yeah, you're right.

14:06 We have a hybrid approach.

14:08 Of course, we build our own models.

14:10 But we want to rely on the best-in-class

14:13 frontier models provided by NVIDIA, such as Nemotron.

14:17 Of course.

14:17 Our optimized model by NVIDIA and available through NIMs,

14:21 which, as I said before, enables seamless deployment.

14:24 It's super easy.

14:26 Or we have also a partnership with other model providers, such as Mistral.

14:32 In fact, we select our models and our partners

14:36 based on the performance of the model,

14:38 of course, but also about the sovereignty and the regulation constraints.

14:43 Because we operate worldwide, We have a customer in all industries

14:49 and many customers in regulated or very sensitive industries.

14:54 So we have to comply with all regulations and all the auditability problems.

14:59 Right, right.

15:02 And so from that, we also want

15:04 to calibrate the model with the customer knowledge.

15:08 So we inject the industry knowledge through fine-tuning or RAG,

15:15 depending on the use case.

15:17 But more generally, we believe in open standards.

15:21 And so we embrace and we support open standards, such as MCP or agent-to-agent.

15:26 In fact, it empowers our agentic platform

15:29 to leverage third-party industrial system and enable,

15:34 in fact, interoperable or cross-system agentic choreographies.

15:40 I wanna ask if we can dig in a little bit to a specific use

15:44 case to kind of get a flavor for some of the things your customers are doing.

15:48 But maybe if there's an example that comes to mind you could speak

15:51 to that really illustrates the use

15:53 of the Virtual Companions and the Dassault platform.

15:57 I think one super cool example, I think, is LEO mechanical designer.

16:06 We showcased this live, this new Virtual Companion,

16:10 in our 3DEXPERIENCE World conference last

16:14 February with Jensen attending this conference.

16:18 And so here, you give Leo a 3D scan or a 2D drawing or a mesh of a part.

16:24 It will activate the industry world model for design,

16:28 orchestrate the AI model and the modeling and simulation solvers,

16:34 and it will perform multi-tier planning, enabling the...

16:39 evaluating, in fact, the mechanical interface of the part,

16:43 find the physics, the kinematics, and the design rules.

16:48 And at the end, it will generate the optimized design.

16:52 It's physically aware, manufacture ready,

16:55 and it will do it right the first time.

17:00 It's a very super example.

17:02 I think it really illustrates our transformation

17:04 from a SaaS to an agent-as-a-service platform.

17:08 And in fact, it...

17:11 In fact, with that, we are giving

17:13 to our millions of designers the power to innovate faster.

17:18 But it's not just about speed, it's about reliability and trust.

17:23 And because you know that your design works because it is born from science,

17:28 from physics, and is augmented with your industry knowledge.

17:33 That change that you referenced

17:35 from a SaaS company to an agent-as-a-service company.

17:39 Kind of from a philosophical standpoint, I guess,

17:44 or an emotional standpoint, does it feel natural?

17:47 Is it a big shift?

17:48 Is it just kind of part of the way

17:51 of doing things to keep innovating and delivering for your customers,

17:54 and so it's just kind of the natural progression of things?

17:57 How do you think about it?

18:01 It's really about, in fact, with the rise of AI, we think ourselves...

18:08 What is the deep impact of AI in what we do, in what we deliver?

18:14 What will be the new experience for the user?

18:16 What will be the new technology?

18:18 We all see the Claude Code, etc.

18:22 What if you apply such transformation to our industrial software, in fact?

18:27 So it came from that, in fact, really.

18:31 And so this is a lot of discussion and brainstorming at Dassault Systèmes.

18:38 And in fact, we don't want to add AI on top of what we do.

18:42 We want to put AI at the core.

18:45 And this is why we are working with NVIDIA on the different topics.

18:48 what's a typical way to get started?

18:50 What's the first project that a customer might typically undertake

18:55 to get started with Virtual Companions and working with them?

18:59 I think...

19:01 You should start from your core business and your core challenge, in fact.

19:07 Right, of course.

19:08 This is where you will have attention from your teams.

19:12 This is where you have your knowledge,

19:14 your deep knowledge and your deep know-how.

19:17 And this is how you know to measure

19:21 the real impact of your AI and agentic transformation.

19:26 And we have an example of connecting to LEO mechanical design.

19:31 We are working with NIAR.

19:34 And NIAR is one of our customers working with us on Virtual Companion.

19:38 And what they are going to do is

19:40 they recreate the Virtual Twin of existing aircraft.

19:45 It means that they are creating thousands

19:47 of parts without access to the original design.

19:51 So basically, they disassemble the aircraft

19:54 and recreate virtually piece by piece.

19:58 So, of course, with LEO, you can imagine how it changed their life,

20:02 automatically generating the 3D part from their multiple sources.

20:07 That's incredible.

20:09 So like everything else in technology, in AI now,

20:13 Virtual Twins, Virtual Companions, simulation, just accelerating,

20:19 advancing so quickly, and obviously agentic frameworks and models are

20:24 developing just as quickly, if not faster.

20:27 What's next?

20:27 What's on the horizon for Dassault Systèmes?

20:30 What are the kinds of things you're thinking about?

20:33 And then if you're game to take it a step further,

20:36 where do you think agentic systems and the idea of virtual coworkers is headed?

20:44 First, I think Dassault Systèmes' strategy is fully

20:46 aligned with the recent NVIDIA announcement about NemoClaw,

20:51 AI-Q, all the the agentic stuff.

20:53 And the rise, in fact, of the long-running autonomous agents.

20:57 And we fully agree on the associated

21:01 industrial challenges— security, compliance, etc.

21:05 And tomorrow, our Virtual Companion—AURA, LEO,

21:07 and MARIE—we believe they will stay awake

21:10 and they will continuously monitor your factory, your project execution,

21:16 your supply chain in real time, and they will proactively optimize it,

21:21 optimize the Virtual Twin without being prompted by a human.

21:26 So it will create, in fact, I think, a closed-loop autonomy.

21:30 And because of our industry world models are grounded in physics,

21:35 I think the agents can use a Virtual Twin as a gym to train themselves.

21:42 So they can run, in fact,

21:44 millions of simulations or design experimentation and present to you,

21:48 to the human, to the engineer, the proven solution.

21:53 And you just have at the end to validate.

21:56 And from that, the Virtual Twin, in fact, becomes a self-evolving asset.

22:01 That gets smarter day after day, in fact.

22:05 Nicolas, there's so much going on.

22:08 For listeners who want to learn more,

22:09 want to learn more about the 3DEXPERIENCE platform,

22:12 about Dassault's work with everything we've talked about,

22:16 Virtual Companions and industry world models, where's a good place to go?

22:20 The Dassault website?

22:22 Social media?

22:23 Are there research papers?

22:25 Where can listeners go to learn more?

22:28 Mainly on the Dassault Systèmes website, 3ds.com, or on our LinkedIn page,

22:32 where we are communicating more and more on AI.

22:36 Thanks also to the NVIDIA collaboration,

22:38 we are posting more and more about what we are doing.

22:43 So, yeah, perfect.

22:45 Yeah, that's free and connect with us.

22:48 Excellent.

22:48 Well, Nicolas, again, congratulations on all the work and thank

22:52 you for the years of collaboration with NVIDIA.

22:54 Thank you.

22:55 And best of luck in everything you're doing.

22:57 Thank you to NVIDIA, to the team, the incredible team.