0:00 I think we’re good, I hope, cool, thank you everyone for being patient. So I’m Alexander Lavin from Pasteur labs, and today we’re going to talk a bit about simulation, and it’s really, I mean, it’s the first thing people see when they go to our website, and they just assume these are the guys simulating everything, that’s kind of the idea, but it’s really a bit more focused, and especially for this crowd we’ll focus a bit on mechanics and engineering and fluid dynamics and additive manufacturing. But at Pasteur Labs we look at the promise of, or this rather expectation of what simulators or digital twins could do, should do, what they’re hyped up to do, a, and it’s really this fallacy or misconception of digital physical parity, like what you do in silico is obviously exactly what happens in, in the physical world. Thankfully this audience, I don’t think I need to convince anyone that’s not the case.

1:04 So Pastor LS, we’re a team of about 25, almost 30, industry hardened research software engineers, what we do is, is we invent, validate, scale new ways of integrating AI and simulation, and in particular in physical sciences for digital engineering. This is really what we call simulation intelligence, or, or SI, and our approach is, I mean, more nuance than this, but it does really differentiate from a lot of deep learning or other machine learning approaches that look at the existing CAE tool chains and think, how can we throw machine learning at the thing, how can we automate, accelerate, predict where the user is going to click next. We asked the question, how can we as humans better understand physics and better understand our world and build interesting things in that world, and AI could be useful to those ends.

2:12 I feel like it’s appropriate to also set some expectations, so what exactly is AI, and this intersection of physics, it’s in the realm of physics, whereas there, I mean, simulation and optimization are pivotal, I mean, understanding and predicting the behavior of complex systems, and artificial intelligence, not necessarily deep learning. So AI, through symbolic or machine learned models, can potentially offer new paradigm in the creation of these simulations, or new ways to approach optimization of physical systems, design of experiment, all the way from microfluidics at the smallest scales to AST physics and everywhere in between.

S scl, or scientific machine learning, is more precisely this merger of computational sciences and datadriven machine learning, it’s the subset of a subset of a field for bringing domain knowledge into physics solvers, or with physics solvers, into machine learn models and data driven models, and deploying those with accelerated computing platforms, and really at Pasteur labs this is our bread and butter. As for these things, I don’t think I need to really clarify how I really feel, we think that there’s a lot of promise for generative AI in user adoption in these tool trains, we think that LLMs can help coordinate or even bring together silo teams across mechanics and aerospace and chemical engineering, those just aren’t necessarily the things that we focus on, they aren’t the things that we, as an early startup, try to build our company around, this is what we do.

4:11 So this is a look at simulation intelligence, I encourage you to look up that paper, it’s long, don’t read it, I’m definitely the only person that’s ever been through all the words in that paper. What we’re going to focus on is how we think about use inspired R&D in physical sciences, and what that means for improving, working with, or really reshaping digital engineering. I promise this is the last time I’m going to make fun of AI.

4:47 So with this crude example, we have a somewhat of a whiteboard sketch of a subsurface modeling problem, you’re building this CO2 injection system, and there’s obviously no like, hey, let’s throw AI at it and vo, but there are ways that we can take AI and ML and think about the many non-trivial, nonlinear steps that are actually in between what we’re sketching on a whiteboard or an engineering pad, and what makes it past manufacturing and scale deployment and testing and has life the real world.

5:29 At pestor lebs we’re building software that is based on differentiable physics, or rather differentiable programming, that is Taylor Made for doing physics, and that enables continuous information to flow between those many steps, the many steps from the whiteboards, the actual thing in reality, and with the continuous information, and in particular information that is machine learnable, it speaks the language of ML, then we can try to get what is designed in silico to match what is tested in situ, and feedback actual, as manufactur data, performance data, to the design engineers in the first place.

6:15 So this is a look at our Auto physics software that aims to orchestrate the complex models and simulation engines that you need to work in advanced engineering and manufacturing these days, and again, what this software does is it provides a userfriendly kind of cool looking interface that doesn’t try to replace your existing tools, it’s not the new ancest, or Autodesk, or anything close to that, but what it does is it packs this automatic differentiation under the hood, and geometric data structures that allow you to just drop your existing models and codes onto the platform, work with a highle user interface, eventually we have an SD K for our power users, as we call them, and it’s end to end differentiable.

7:09 Now one of the ways that we talk about Auto physics, u, and the ways that we kind of describe different users and their perspectives, and how they will want to kind of, I don’t know, fast forward through the complexities of things, but really zero in and double click on the things that they’re experts in, we like to think that Auto physics could bring this multiplayer mode and dynamic digital engineering, instead of the static tool chains and the lone mechanical engineer sitting at their workstation passing off three simulated data points to some manufacturing engineer that they never meet, let alone get feedback from.

7:51 So for now we want to focus a bit on those interfaces in Auto physics, and the information exchange that needs to happen, and in particular with CAD and CFD challenges, when you have really, I mean, beyond hello world or anything of interesting complexity, the multiple components modeled with different physics solvers, or you have joint optimization over multiscale systems, even systems that vary on different temporal scales, machine learning offers promise, it can be promising, but it can’t talk to CAD or CFD, it doesn’t speak the same language. If you think about doing a PhD and building one tool chain, you’ll get one experiment, and good luck trying to get that feedback loop and anything that could resemble an automation loop for optimizing CAD and CFD with downstream data or quantity of interest sort of objectives.

8:54 So speaking of users, we have, I mean, I don’t know, ways that we describe, des cribe our users, they wear different hats, and it allows us to really think from, I mean, day one, what that multiplayer mode, how that can actually, you know, emerge in the not too distant future, and force us not to be so constrained and focused on, like, the cool ML SDK or some rest endpoint that’s perfect for this web app deployment, but really we need to involve and orchestrate all of these perspectives. And like anything machine learning in AI, it really, it really starts with data, and the data availability, it’s something that, u, whether it’s partners we work with, or, or customers, or even just talking out in the hallways of this conference, people say, we have data, or a client has the data, it really signals to me they haven’t, they haven’t tried it yet. If you don’t think data is a problem, you haven’t really approached your data problem.

10:10 So these are a few ways that we kind of have these buckets for sort of value ad of the Auto physics platform, and how we can enable CAE experts to not build this on their own, not worry what machine learning ready data actually means, but plug and play, and that means parameterization to work with machine learning models and, and training and inference algorithms, it means vectorized, or data structures and formats that can actually interface with downstream machine learning libraries, and always an afterthought, but to do anything interesting it means scale, and scale for software engineers, I mean that means continuous integration, continuous deployment, it means regression testing, nightly builds, and these are things that have to be from day one in your data pipeline.

11:12 This is, with all these slides I don’t want to, I feel like I should thank my team for all the things, this one especially, just knowing the engineering lift that goes into a concept of taking your pick, or mixing and matching a tool chain of CAD, CFD, FEA, CAM, and pressing a button, it’s a little more complex than that, but pressing a button and getting these CA data sets out that are machine learning ready.

11:41 Here’s a little demo, it’s recorded, we’re not allowed to do live demos, of essentially a CFD expert, doesn’t know any machine learning things, would be scared if there’s any machine learning architectures to sort through, but is going through, picking their favorite set of tools, and the tool chain selection, and the search of what sort of machine learning engine, or accelerator, or machine learning training behind the scenes, it’s all behind the scenes. So as it’s going again, and the user eventually gets that pink block in the middle, you’ll see they trigger this execution of a, a data set, and checking the logs, seeing that is running, and then is able to come back to the platform and keep building, it takes, sure it takes a while, but coming back here and actually setting up another set of our products on our platform for deployment.

12:43 So data sets here, they aren’t CSV files, they’re not what you just export from existing CFD tools, you export a software container, you export directly to an S3 bucket, these are things that software engineers has been doing, I mean, decades, and it’s overdue, let’s say, for mechanical design engineers. It’s really cool when you think about the variation that you can get when you aren’t that loan engineer kind of iterating over a few parameters and then sitting and waiting 4 hours, 12 hours, often more, to generate one data set, and then think of, like, all right, how do I want to tweak this. Now with the automation in the scaling, and pressing a button, you can, as a mechanical engineer, and your programming is like, I’ve done some bash, sometimes you can actually deploy to your infrastructure, your cloud, so you don’t have to try to, I don’t know, think about what it’s like paralyzing across, I don’t know, 64 A100 GPUs, or what sort of memory you need on your CPU on that machine, especially if it’s mult machines, and the automation gives you really rich variation.

14:06 Can we see the videos? Oh, cool, I don’t see the videos on M, so just looking at a few of these, and we see that there’s very different behavior, let’s see, in temperature variation, in velocity of these particles, when you look at another instance, you have the effects like, that’re just boundary conditions, so you have sort of your, your, your normal behavior, but when you’re generating all of these automatically, and you have algorithms that are intentionally searching the, the space of parameterizations, you get a really rich distribution, and things that you as an engineer maybe would not think to model, you get instances that are not always guaranteed, but there’s always those that happen in the real world that you never think of in simulation.

14:59 And that’s why, I mean, pick a industry or a sector, and no one trusts what you do in silico is going to actually translate to the real world. If we can simulate the things, we can try to simulate all the things, we can get closer, and going back to the team, a lot of us come from AIML backgrounds, and know that, like, the workhorses behind AIML, that’s automatic differentiation, and that’s data as a first class citizen, have been the motivation and the perspective with which we tackle problems like automating Anis and Autodesk and other tools, u, and how we build that in a future proof way, so when our platform does come out, in a few months, it can be extensible for your custom solvers, your, your whatever infrastructure you have on prem.

15:51 And here’s just a snapshot, this is from robotics data set generation, gosh, 10 years ago now, with a setup that actually looks a lot like the imaging display out in the hall. Now to what the talk is actually titled and supposed to be about, differentiable physics, programming doesn’t happen unless we have the data that is machine learning ready in the first place.

16:24 Now let’s, going back to this sketch of the system of systems problem, and our Auto physics product, what we’re building in Auto physics, and when I introduced this company as inventing and validating and scaling, I mean that validating part is, I don’t know if we care to track time, or if I ask my team what are we doing every day, it’s so much of this actual validation, because we want to provide a catalog of template solutions and demos for those templates in different domains. So if you’re coming from a subsurface energy, like the CO2 injection, like that’s your, your field, you would be able to find a domain or something close to it, and find your way into the catalog of these surrogate based Auto physics solutions.

17:13 Or if you are coming as a, an M&S solutions engineer, or a machine learn engineer, and you’re thinking multi-objective optimization or quantity of interest over the system of systems, you want to do active learning, you want to do data simulation or calibration with your as manufactured data, you would be able to find the template to start, you aren’t starting from scratch, you aren’t asked to do any of the machine learning things, that’s our specialty, but we also, if you want, you can tap in that SDK.

17:47 When we come out with it, when we talk about these things and actually describe what, what is a test erect, it’s been cool to hear from a few partners and customers that their first impression is, it’s black magic, but then they get into demos and stepping through some of the docs, and like the tutorials that come with the catalog, and it is a necessary component or, or product of not just our platform, but how we envision that simulation intelligence ecosystem. Really that ecosystem is one that we think needs to be driven by digital engineering and, and the manufacturing, u, and design problems here, and ideally is one that can really level up all the things that we’re doing and trying to build in the real world.

18:46 So with that, if you want to find out more, reach out, grab myself in the hall, some of the team is here, promise we’re friendly, and we’ll share more insights than I showed here. Thank you.