Hello everyone. Welcome back from lunch. I’m Mike Park and I’m here with Flex Compute talking about many common thing themes that you’ve we’ve seen through these last two days. So to begin with this is sort of a cartoon to kind of indicate you know the the multi-disiplinary design that’s sort of desirable. We’ve seen this from from a number of participants. You know in the center is an aircraft in this case but but it could be electronic device you know many other things it’s standing in for the I’m kind of portraying the geometry as being the center of this maybe that’s just from my my background u but you would imagine this geometry held in some kind of product life cycle management system and from there you need to historically manually post pre-process process or prepare this geometry for different disciplines.

0:58 And as the longer this round trip takes, it adds latency, it adds you know you know slows down the process, adds impedance and then at certain point you know some of these disciplines are working on different revisions if if this thing is is evolving and this becomes really a roadblock for doing multi-disiplinary design. So and this sort of is again another cartoon to kind of indicate you know historically this this process can take on the order of days or or longer in order to prepare the geometry.

1:57 There’s been you know a significant investment in meshing and solving technology over the years but this geometry preparation has been sort of a you know a roadblock. And by addressing it there’s a there’s a potential for higher throughput through the process and also this opens up the potential for automation which allows for not only you know engineers to become more effective but also agents. So by removing these roadblocks we’re able to start with this geometry import and then do cleanup to it.

2:42 And this is we think one of the really key parts to enabling this whole process cuz cuz once we have sort of a a usable representation of the geometry, things like meshing become a lot more straightforward. And then adding in other aspects of AI, we’re able to do solver setup in a lot more robust and fast way. We’ve had a a long investment in GPU solvers. So that part has always been fast, but now the rest of the pipeline is keeping up with it.

3:16 And then now this all falls downstream to the the insights that we’re able to get. This includes both training and inference from surrogates as well as other multi-disiplinary design. So this framework we’ve lined out is really kind of underpinned by a unified py python API and this is was originally designed for humans obviously but now as the rise of agents it’s becoming obviously an equally important thing and one of the things we’re sort of discovering is these are not competing interests that a a well-designed API both suits humans and agents and the agent agents or the the agents typically interact with it through the API but users also have the ability of using things like web UIs and and other sort of visual tools where actually in some places the agents want to have this as well.

4:16 This this this API layer can call into what we’re calling numerical physics or kind of the classic solvers in the in the different disciplines that produce verifiable results as well as neural physics which underly you know foundational models surrogates and and other other products in that that category and then these interact we have both training and inference back and forth Earth. They can also support modeling within these these classic techniques.

4:52 And then underlying this is the geometry layer that allows you know import from kind of classic BREP as well as discrete forms of the geometry. So some of the key aspects that we’ve built into our geometry ingestion pipeline is it’s high fidelity where the CAD is well posed and robust to imperfections that typically hamper analysis. It enables automated healing by being aware of the topology of the problem.

So then tasks like ceiling transformations and boolean operations are are possible. It’s made reliable for meshing by introducing strong mathematical guarantees in both water tightness and manifoldness for consistent reproducible results. It’s universal in that it can take both boundary representation and discrete models to to into this representation and then it’s agnostic in its outputs to support different disciplines. So here is sort of sort of a cartoon for classic challenges that you would run into.

6:08 This is an automotive case, but I think it’s common with a lot of other disciplines. In this case, the the car has panel gaps which are there. They’re part of the design. In order for the door to open and close, you have to have these gaps. If you’re doing serial aerodynamics, you may not want to resolve those. And then there’s also maybe defects in the model. In this case, the grill is intersecting part of the radiator.

So you may have both these sort of design challenges or designed in challenges for analysis as well as basically defects or errors in the model. So we’re able to look at this entire model in its entirety, do classification and then be able to extract out certain aspects of the model for and this is based on sort of the type of analysis you would want to do. For example, if you wanted to extract the outer mold line for external aerodynamics, that’s possible.

7:09 It’s also possible to extract the inner mold line, for example, if you want to do acoustics. So and this applies to you know other other disciplines if you want to do internal flow or or some of these other cooling problems that the the ability of doing this classification extracting out the part of the geometry that that’s necessary for your analysis. And so here’s here’s another example more in sort of the aerospace field.

7:38 This is there’s a maul of a rocket engine. This is also full of defects. These green areas have been categorized as being internal to the model. So you can see there’s places where bits of the piping intersects with other parts. So this is not a watertight and there’s also baffle surfaces in here. And then this is again through classification we’re able to extract what the outer mold line is and then mesh it appropriately.

8:10 And then this is also you know goes to volume meshing and then you know we’re able to do CFD on this. So that was for a you know sort of sort of a you know for steady RANs. We also have capabilities in terms of hybrid RANs. Oh it’s not playing. We have capabilities in terms of hybrid rans. So and in in this video you would see the not only are the rotors rotating and having these these weights that descend but the entire vehicle is approaching the ground.

8:57 And this was a particular analysis that was done that was found you couldn’t do this in a quasi steady sense that it was actually the descent ended up impacting the design. So having this level of fidelity is really important to be able to accurately represent the physics that a vehicle is going to experience. And our underlying GPU flow solver is not only fast as sort of demonstrated in in these complex cases but has also been fully verified in public workshops as well as benchmark cases.

9:35 This was the video I was trying to show previously about the vehicle and it’s descending. So as the these rotors are spinning they’re producing wakes that interact with the ground and that interaction is actually significant and as well as the descent of the vehicle. So now this is another interesting case in a very different discipline. So this is two spacecraft approaching each other for docking where one vehicle is coming to service the other vehicle and you can see it’s firing off reaction control jets as it approaches one vehicle which may be cooperative or may be inert you know if it’s coming to be serviced.

10:13 And you’ll see here in a moment that we’re resolving the pressure field that’s on this vehicle. And these pressures obviously produce forces and moments and there’s high fidelity but filling out an entire simulation database for this requires months. So that this is a place where we were able to use surrogating to build a surrogate that reduced these highfidelity evaluations to 100th the number that they needed previously.

10:46 So this this takes this month-long preparation for a mission down to days. This is a highly multi-dimensional space that this this vehicle is flying through with protected with respect to position and orientation. On the left you’ll see a 2D slice through this highfidelity data. And this is sort of a showing what’s happening with the this adaptive training technique we’re doing. So first we start with a random sampling of points of the entire space and then an uncertainty quantification runs and as the uncertain parts of the the this large dimensional space are identified additional highfidelity evaluations are called training redone and the the uncertainty is again quantified and then from this we able we produce and you know an adaptive result that allows for a massive reduction in the the high fidelity evaluations.

11:55 So here is another interesting challenge that originally started from the human world and now I think is going to become even more acute when it comes to enent engineering and that’s if you if you ever had the situation where you’re doing a design you have artifacts that you know in this case we’re we’re saying a plot is like one of these artifacts. So, in order to make a design decision, you you’ve made some artifact that that educates that design that you you know, you have a plot that says my lift has to be this or my drag has to be this or my weight has to be this in order to close my design.

12:32 These artifacts tend to be static and then you don’t know what was used to produce them. In some cases, you may not know who produced them or that person may have left the company, you know, as as time goes on or older projects. So having the ability of having these artifacts be queryable that shows the graph of information how it flows into it. And then once you have that you can also ask some really interesting questions where if you have an artifact you can look at aspects of it and you can ask an agent to tell it about things.

13:09 You you can ask it like for example you know what mesh was used to make this point that became a plot. What what you know what person or what agent. So this is going to become I think in in increasingly important as we move in an environment where we’re co-working with agents or until they take over or you know where where humans fit in this mix that you need to answer the question of where where where where do we where did this data come from and how is it going to impact my design?

13:44 And and again this is one artifact a plot but you can imagine there’s other artifacts you know a geometry could be an artifact that’s attributed and we have ways where it obviously works best within our ecosystem but we have ways of having links and sort of tags to be able to keep track of information when it leaves and re-enters. So now we we’ve seen this sort of image or I should say this sort of use case before because we have a wellposed AI that sort of on on on day one we were able to hand it to an agent and it was able to read the docs and form scripts in order to you know run our application and and do you know do analysis on our our behalf.

14:31 We’re continuing to improve this process and streamline it. We’re learning as much as everyone else is about how to do this effectively. But one of the interesting things is with this sort of process is it’s giving us sort of a a way of improving our product not only for agents but also for humans because we’re finding that where we improve our API or where we learn where the friction points are with agents those often overlap with where humans have mistakes setting up a problem or or or any of these other challenges.

15:06 So we now have sort of automated tools of not only making our product better but also making our interface better. And in you know in sort of closing we’ve got an amazing group of people that we’ve collected to work together we love hard problems and we’ve made custom solutions for a number of customers that have come to us and had needs. So it’s been you know it’s it’s been a really fun it is a very fun you know thing to work to work here and all the the fun challenges we get to.

15:42 So in closing this is you know we we sort of brought to brought full circle this end toend workflow which really starts with geometry and then ends in this insight and enables both humans and agents. And if you have any questions you can reach me at mikeflex.com. So thank you.