The Unreasonable Effectiveness of Simulation Intelligence
Interview with Alexander Lavin – Pasteur Labs
Ahead of his CDFAM 2025 NYC presentation, we interview Alexander Lavin, founder of Pasteur Labs, who’s pushing the boundaries of simulation intelligence for real-world engineering.
Building on his presentation and interview from last year, on data bottlenecks in Realizing Differentiable Physics in Digital Engineering, this discussion dives into the evolution of SI, the “simulator vs simulation” paradigm, and the practical impact Pasteur’s innovations can have on engineering workflows.
Pasetur Labs will also be hosting feedback sessions at CDFAM, inviting practicing engineers to help define the future of digital simulation itself.
Your upcoming talk is titled “The Unreasonable Effectiveness of Simulation Intelligence.” which has a beautiful poetry about it, that hints at a 1960’s Winger article on how some pretty basic observations in Physics in the past have led to an outsized understanding in a broad range of applications. How do you propose that simulation intelligence might have a similar impact?
Wigner put forth an important, and yes beautiful, truth of the world: mathematics often precedes its scientific use, suggesting not just a relation but rather this uncanny preadaptation between abstract formalism and empirical truth.
Maybe 15 years ago Terry Sejnowski—whom I was fortunate to meet briefly back when I worked on neocortical algorithms—analyzed the deep learning field from a similar perspective, making the case that empirical success of deep learning, even back then, is surprising or “unreasonable” given our limited theoretical understanding.
Now SI—that is, simulation intelligence, as declared in our 2021 paper—is largely about the combinatorial possibilities available when SI motifs, like causal reasoning and multiphysics modeling, are brought together.
One view of SI’s unreasonable effectiveness is “1+1=3”, like how surrogate modeling & probabilistic programming in synergy can unlock capabilities beyond the individual components, such as modular multiphysics with uncertainty-aware models.
Another view of SI’s unreasonable effectiveness, one that is more tangible for users of our technologies, is the compounding value of SI-driven applications.
Consider an engineering simulation produced with traditional CAE tools: it’s a one-off, for that one task or giving you one data point, it’s static.
Our SI software produces simulators, which continue to update & validate for the next 1000 tasks well beyond that first one, not to mention flexibly coordinate with other simulators—it’s dynamic, at the speed of need, like engineering teams need to be.
One of the quotes from your NYC talk that resonated with a lot of people was: “If you don’t think data is a problem, you haven’t really approached your data problem.” Could you expand on what that means in the context of simulation, ML, and digital engineering for those who still think there is not a problem?
Engineers spend weeks putting together each CAE simulation, which then take days or even weeks to run. What industry is now starting to realize, what I elucidated in the talk last year, is that it’s non-trivial and arduous to transform those CAE simulations into datasets usable for AI or ML—even if your team has the Data Scientists or ML Engineers capable of such data engineering & MLOps, you’re looking at another 1-2 months of dev work before you can consider plugging in AI tools or building surrogate models.
We care about the end-to-end engineering pipeline, which means building capabilities on our platform that make this data engineering easier for end-users, removing this massive barrier to AI adoption—amongst others like model deployment and interoperability.
What do you intend to present at CDFAM 2025 in NYC? Will we see more of AutoPhysics in action (people love to see animated gradient maps), and how does this build on what you shared previously?
For sure, presentations and demos tend to be exciting when you’re building simulators in jet propulsion, nuclear fusion, and so on. You can expect physics-AI insights along with demos running Pasteur’s AutoPhysics product in multiple domains, with an emphasis on the “simulator versus simulation” distinction I alluded to earlier: end-to-end solutions that can continuously improve, adapt to shifting environments, extend to additional tasks, and thus generate more insights beyond the initial use-case, or what customers can expect with our SI Platform versus the one-off solution or bespoke model they get from other providers.
Who would be an ideal early partner or customer for Pasteur Labs, and what would make that collaboration successful?
We’re looking for companies who want to push the boundaries of what can be accomplished in simulation, the impact of which is not incremental rather it is impactful up-and-down their engineering roadmaps: whether that’s accelerating their design cycles constrained by simulation dev- and runtime, integrating ML into their workflows to get the most of their data and their expert knowledge, or moving their expensive real-world tests into software because they can rely on simulation intelligence testbeds that minimize their sim-to-real errors.
We want partners who see the potential of AI-powered simulation workflows to reshape engineering, and are excited to help define what that looks like in practice. What makes these collaborations successful is mutual learning.
We bring AI-native physics simulators, CAE data engines, orchestrators & pipelines for accelerated & distributed computing (our open-source “Tesseract” system), and soon an SDK.
They bring deep domain expertise and real-world constraints. Together, we figure out what actually moves the needle and solve real-world problems—not what sounds good in theory or looks flashy on social media, but what genuinely changes how engineers work.
The fast track to working with us as a customer or partner is to start the conversation today: hello@pasteurlabs.ai
Finally, what are you hoping to gain from your time at CDFAM? Are there particular questions you’re exploring or people you’re hoping to connect with this year?
We’re excited to engage with the CDFAM community—the signal-to-noise ratio here is exceptional.
We’ll have Pasteurians from our NYC office and Europe locations at the event, eager to show you the products we’re building and, more importantly, to learn from you.
This year we’ll be running several SI Platform user experience sessions in coordination with the event, to dive deep with CAE Data Scientists and Computational Engineers who would like to share their insights and work with Pasteur Labs experts—this is your chance to shape what gets built next!
Your frontline experience—the friction points, the workarounds, the “I wish this worked differently” moments—will directly influence the next generation of digital engineering, namely the “IDE for reality” we’re building at Pasteur Labs.
Interested engineers can find out more and sign-up here.
Attend CDFAM in NYC to connect with Alexander and other leading experts driving the adoption of AI and ML in design engineering and architecture, spanning academia, industry, software development, and practical application.
