CDFAM NYC 2023 · New York · 14–15 June 2023

Robust Geometry Processing for Physical Simulation and Shape Optimization

Listen as a podcast

Abstract

This recording is part of the CDFAM Computational Design (+DfAM) Symposium and features Daniele Panozzo, associate professor at the Courant Institute of Mathematical Sciences at New York University.

Presentation Title: Robust Geometry Processing for Physical Simulation and Shape Optimization

Daniele Panozzo presents an in-depth session on ‘Robust Geometry Processing for Physical Simulation and Shape Optimization’. Panozzo, who heads a research lab at NYU, has been instrumental in advancing meshing techniques that enable automation and optimization in design. In this presentation, he elucidates these advancements, with a particular focus on how they have been made available as open source code on GitHub, benefitting software companies like nTopology. He specifically discusses the influence and applications of projects like TetWild and PolyFEM. These innovations underscore the future potential and scalability of robust geometry processing.

This session is intended for professionals, researchers, and academics in the fields of computational design, additive manufacturing, and geometry processing. Those who are interested in the impact of robust geometry processing on physical simulations, shape optimization, and meshing techniques will find this session particularly useful.

Transcript

From YouTube’s automatic captions, lightly cleaned; expect some errors. Each timestamp opens the video at that moment.

Read the full transcript · 4,479 words

0:08 Hello everyone, thank you so much for inviting me here. I’m a associate professor at NYU and I lead a group there working on geometric computing. So today I would like to start a talk by with this definition of catastrophic failure, right? A catastrophic failure is a sudden and total failure from which recovery is impossible, like, I mean, you have a puncture on your tire of your car, it doesn’t recover, you have to stop, some human has to be there and fix it. One of the things that I found curious is that while most of us, whenever we design something, we are very very careful about trying to minimize a potential catastrophic failure of a part or avoid it entirely if we can. Somehow, especially in this community, there is a very very high tolerance to catastrophic failures in the design tools, right? I mean, you take, you want to take a cut model and you want to solve an elasticity problem on it, good luck, right? That will require cleaning and quite some work.

1:12 So the goal of my group, I’m lucky to have this group of people working on this problem, the Geometric Computing Lab, it’s in downtown Manhattan, is called by me and Professor Dennis Zorin. We are all working toward finding ways to not have catastrophic failures in geometric problems, so the the entire goal of the lab is to figure out ways to do design, simulation and optimization of physical systems. But before getting there, just a quick overview of why are we doing it. So here is an example of mechanisms, this is an object downloaded from Thingiverse and tested in our simulator just to see if it works or not. So a simulation is useful when it’s able to predict a system, like in this case where we have a high speed ball bouncing on on a wall, this is a soft golf ball made of foam, so that’s why it deforms so much. And the main reason that I have for trying to simulate physical systems is because it’s much cheaper and faster than actually doing a real experiment.

2:24 So a physical simulation in this setting, it’s useful, or it’s interesting, it’s valuable if it saves time in a design iteration, and in particular there are two features of a simulation that are crucial here. First of all, it needs to be predictive of what you are modeling, because otherwise there is no point in doing it, and the second thing is that it needs to be efficient. But I would like to propose a definition of efficient which is a bit different from what it’s usually discussed and what has been discussed in many other talks in in this Workshop. I don’t think it’s so crucial to be efficient in terms of CPU time, CPU are pretty cheap, you can have many in in the cloud to run something in parallel, but one thing that is very precious is time, human time. And so what I think it’s important here is to optimize, trying to minimize human time more than than machine time.

3:18 Just going back, so it’s a long time that as humans we are looking at methods for doing simulation, and the first one that I actually found, you can think of it as a sort of analog computer, is the Antiguitera mechanism. If you never saw this, there is a very nice documentary on Netflix on this, it has very recently been discovered and it’s partially understood how it works, unfortunately not completely because half of it is missing, but it was used to predict astrological positions and eclipses way in advance, and this was a long time ago. Most simulations are done on digital computers, the first one that was publicly available and very heavily used during the war was the ENIAC, which was mostly used for for military applications. But let’s go down many many years in the future, let’s get to today and see what we can do if we want to simulate something, what kind of tools do we have.

4:13 So we can take an existing simulator, we can model the scene that we care about, for example let’s say that we have a box fixed, a box with a load and some deformable boxes in between. This looks like an easy case but it’s really not, because there are multiple components the interact through contact, there is large deformation, there is no linear elasticity, and unfortunately to get something like this to run it takes quite some effort and a lot of expertise in how the simulation methods actually work. You need to fix a mesh, you might have to fix parameters for contact, you might have to fix the time step, and after many many iteration you can get it to work for some set of parameters. The tools here were really not designed well, they’re not prepared for design purposes, they, these tools are excellent for taking an existing design and validating it, but if you want to do a new design you probably don’t need to simulate one thing, you need to simulate a million things.

5:09 And so the goal here, I think these kind of tools are really not the right ones, and for having some tools that allow supporting this faster design iteration it really requires to rethink it more or less from scratch. So this is what we are working on, and I would like to explain like what is my vision on these, and then tell you where we are in this journey of algorithmically rethinking visual pipeline. So what I would like to have is something where you give me a geometry base, which is outside from inside of a object, I have some boundary conditions, someone is sitting on this chair, and then ideally nothing else. I shouldn’t bother about creating a mesh or not creating a mesh, what simulator do I use, what parameters. What I really should provide is what is the final error that I want to have on my stress, and when the rest will be automatic.

6:04 There should be no reason for a human to actually be involved in this, and if I see that this deforms too much I can just go back manually and do some changes and try it again, but if it’s all automated this will be like such a relief, like this is such a better way to to proceed. But this year is kind of trivial, so what happens if I go for something bigger, like for example these two design that are coming from anthropology? Well, I argue that nothing should change, there is absolutely no difference, the fact that this is much more complicated, it’s a technicality that should not affect the algorithm or the parameters that you need to tune. And what happens if you have contact? Well, you should say yes, I do want to model contact, but it should stop there, right? The rest is just physics.

6:51 But you just work out here, just to give you an idea of a scale, we have a simulation of this soft ball using non-linear elasticity bouncing on this wall. We have 2 million tetradra, around 5 000 collisions at every time step, and what is crucial here is that the code has to be provably correct, because otherwise at some point it’s going to fail and tuning parameters here is way too hard. And what if I have friction? Well, similarly, you can enable friction, if you measure reflection coefficient between your surfaces you can simulate, but then when you see something like this here you might notice that looks a bit strange, right? There is some sort of strange non-smooth behavior in the middle, and one might wonder, is it up artifact of a simulation or not, and it turns out it’s not.

7:38 This is what is usually called stiction, or a transition between static and dynamic friction, this is very typical if your filament got stuck in a 3D printer, you might have dealt with this, but this is something natural and important to model if you care about correctly representing the physical system. So all of this is about forward simulation. So like now, for a second, let’s assume that we have this, we have this magic ballet that allows us to do forward simulation. So what we can do is take in a domain and boundary conditions, we can simulate, for example, to compute displacement, and from displacement we can get stress. Here you see an assembly of four components that are interacting to contact and friction, they are not welded together, there is some load applied and there is some stress there.

8:23 So now one thing that you might want to do is to figure out a way to optimize this piece to minimize stress. But instead of doing it manually it will be much easier if you could just say, well, I have a stress field defined on the shape, I want to maximize strength, I can model it as minimizing max stress, I can smooth it out a little bit just to make it easier. But if I could do this in a completely automated way, where you say my design goal is to minimize stress while staying close to these, and what I’m trying to do is to change, for example, the shape of this object. So note that I’m not talking about topology optimization, where we are deciding where there is material and where it’s not, here I’m really talking about let’s slightly change this design by deforming its boundaries so that stress doesn’t concentrate.

9:13 So this kind of problem are usually called PD constrained optimization problem, because you’re minimizing some functional that is constrained by a physical system modeled as a PDE. So here is an example of what happens for this bracket, and you instantly get something that is more organically looking, because this is what tends to like corners, tends to attract stress. But note that there is no extra data provided here, we’re just like trying to find the shape that minimize stresses, and you see that there is still like contact between all these pieces, there is friction keeping it together in a highly non-trivial way. This really has practical implications, this is just a small experiment just showing that for these clamp you can very easily dramatically increase its effectiveness with small changes in shape.

10:03 And you, if you have this, you can actually push it forward. So you can think about, let’s design families of metamaterials where I randomly start from different tiles, here we start from a few hundred different tiles, and then for each one of them I use this differentiable mechanism to find variations of their parameter of the material optimization, where you actually replace in the material the geometry with one of these styles, so that you can have a single base material that are specially varying properties. And when you, this is something that we heavily work on, and this is actually what motivated us for going for this full automation, because building this kind of family requires running millions of simulations, and so if a pipeline is not completely automated, meshing and PDE solver, it’s impossible to do this sort of analysis.

10:53 This is some example, I also have a printed copy here if anyone wants to take a look at it later, of using geometry to create negative concentration, so there is a mechanism that like naturally appears in here, and note that for something like these where there is a huge deformation you cannot just use linear elasticity. So what is required here is to actually model the shape in the tiles using a proper non-linear material model, for example. We do this sort of work both in 2D and in 3D, and here are some example of a design that we can get. This is a batch stretch test of our pipeline where we take random geometry, random card models, and when we optimize them to get them to bend in a controlled way when a certain load is applied.

11:35 And the core thing here is that you can provide whatever geometry you like, and the algorithm is gonna successfully tile it and get something which is close to the prescribed deformation. The fact that we can do this kind of optimization with no linear elasticity, with contact, and with frictions, open the door to a lot of interesting analysis. Here is, for example, an analysis of a bi-stable lattice, this is simulated at full scale, so this is a single mesh of a full lattice, and this is a bi-stable structure, meaning that it has multiple, every tile as multiple rest states. In the one on the left you see a design where the block is completely released, in the one on the right you see another example where the energy is not fully released, so the structure stays locked, but this is a kind of kind of an interesting shock absorber, because you could reopen it back again to go back to what it was originally.

12:28 And all this sort of analysis are impossible if you restrict yourself to linear elasticity. So the goal of the research is really to get to a point where all of these is fully automated, and we are pretty close, at least for the example that I I showed you, and our goal in general is to get to a point where we want to minimize time, even an error bound, and all the technicalities involved in the solve, including match generation, transfer boundary condition and so on, are fully automated. So one comment is that this is a very hard problem, I believe it’s a very hard problem, it involves computer science, physics, numerical methods, and one of a sort of frustrating thing is that a lot of these problems, both in industry and academia, are actually considered solved, like meshavi that people look at for such a long time.

13:15 But actually there is a huge difference between sometimes being able to get a mesh or having an algorithm that no matter what is going to give you a mesh that you can use for finite elements. So I think there is a lot of space for interesting problems to look at there, and maybe it’s a bit of an extreme view, but my argument is that any algorithm that is not fully robust is not so interesting, because you cannot use it as a building block for other things without having humans in the loop. So in the group we work on these, we spend a huge amount of effort figuring out what are the best ways to discretize space, and we converge to a solution that I think is quite popular in this community, which used to simply use tetrahydra with quadratic bases and maybe quadratic geometric maps.

13:56 We found that this is the most efficient, and given our technology for doing meshing this is actually something that you can use even for very complex geometries. At this point, when when you decide how to discretize space and time, the question is how do you do it, what are the algorithms for discretizing space, how do you transfer boundary condition from your surfaces to the volume, how do you deal with contacts if they appear, how do you solve the actual PDEs, not caring if some elements are of low quality, it’s actually possible to do this ignoring mesh quality. And then what can you do with it? I think PD constraint optimization is very exciting, especially for the design of metamaterials, how to build a curved geometry so that you can have reduced models that can be solved more efficiently.

14:43 And then other applications, we are working on biomechanics and in a few other spaces, so I cannot tell you about all of it in the remaining seven eight minutes, but I want to talk about two things that I think are of interest. One is meshing, and it looks fine, I mean everyone can can tell what this is, but if we look carefully in this spot we see that maybe something is wrong there, and if we look a bit more it’s starting to look bad, but if we really zoom in we see that something really really went wrong there. So now if you are here and you need to simulate this thing for whatever reason, what do you do? You have two options, right, either you say no, this is crap, I don’t want to deal with it, which is what it’s usually done, most software will just tell you, you can’t get a mesh out of this, it doesn’t clearly divide inside from outside.

15:33 Or you can try to deal with it, but this requires completely different algorithmic support for this, because this is not anymore a question of meshing inside and outside, it’s something else. So I argue that we need to deal with this, and one of the reason is because unfortunately we are stuck with a way more common way of representing CAD models, which is by using collection of B splines of nerves, and this is a really bad idea, because usually you intersect nerves you get nerves of higher degree, but then instead of dealing with the IR degree you just close your eyes, approximate with a low order, and things are gonna go bad, and we are going to systematically go bad every time. So the large majority of the models are broken, and it’s very hard to do anything else if we want to stick with with cut, which I think it’s it has to be the case.

16:21 So how common is this? So if you take 10 000 models and you try to match them, how lucky are we? And unfortunately it’s pretty bad, so most of the time it doesn’t work, and work means I get a mesh, like forget about quality. And and the reason for this is not because these arguments are broken, is because the requirement on what the input of the argument is is not realistic, it’s not what it is, a CAD model is not closed, it’s not for yourself intersections. So, and if you want to automate anything with a failure rate of 80 of 20, there is no chance. So our take is that you could try to not use a mesh, but unfortunately this comes with a lot of other issues.

17:00 So what we do instead is we don’t think of the input as fixed, we think of the input as a suggestion, we want to have a mesh which is close to the input, 0.1 millimeters from the input, and by doing these we actually can get a new series of alternate which is fully robust. A fully robust meaning that in the last few years I I’m not aware of any failure of this software that has been online, it has been used by a lot of people, including being distributed as part of anthropology. So here’s an example, you can provide an input triangle soup, and this software will clean it up, giving you a clean surface and a clean volume mesh. It works on very large models, and we tried to break it as much as we could, but there is no failure for these, because most of the computation is actually not done in floating point, and I’m happy to tell you more about this separately.

17:54 Another thing which is problematic is any kind of simulation with contact. Here is an example of a representative case where I’m just trying to stack a bunch of cubes, and depending on the time step that I pick for my simulation they might collapse, they might have intersection, or it might not work at all. This is actually a similar story, so if you take existing software, every column here is an existing family of methods, and every row here is a different simple unit test, mostly like dropping a tetrahedron to another one or a cuban to another one, you see that depending on the parameter choice we are going to either succeed or fail, which is fine if you’re trying to do one sequence, but it’s absolutely not fine if you’re trying to do shape exploration, because any failure will just stop your optimization.

18:37 So also here we work on trying to have a solution which is very different from the entire way context has been handled. We don’t deal with contact after we are introduced, we actually prevent them from happening, and I’m happy to tell you more about this separately, but what you get is you get a simulation method where it doesn’t fail, it’s not possible for this method to actually introduce self-intersections, and this can be used for simulating very high speed scenarios. Here you see a bouncing of this ball with a shock wave propagating, which is very close to what we recorded, this is another large example that I showed before, and of course all of these comes with the cost of extra CPU time, but I think the trade-off of human time for CPU time in tuning parameters is a very good way to make progress.

19:28 Here is another example, you can try whatever simulator you have, either the formable or rigid, if you try it on this example the only question is how many links do you need before it breaks. And in our fibrillation is quite natural, is that you don’t need to commit to how many degrees of freedom do you need, you can start with a course mesh, and then the algorithm can only fly figure out when the mesh needs to be refined to be able to capture the physics, in this case the physics of the impact. And this is something that we we are still working on, this is very new, it’s going to be presented with summer.c graph, but it’s kind of mesmerizing to look at this, these sequences where you see that like the mesh is being refined wherever you see high stress.

20:07 I stress, due to either elasticity or contact or friction or a combination of them, and the model is refined to to better capture the physical system. If you are interested in these, there are many more sequences and the description of how these methods work on my website. Here we made it as soft as possible to just make the point that this doesn’t break, here you’re dropping a ball into a bed of spikes, and even there you get the correct refinement, and you get like a proper capture of these high energy impact in when the spike hits the deformable ball, which you completely lose if you don’t refine.

20:55 So one one thing that I strongly believe in is that for many of these algorithmic parts is really crucial to build this as a community, more than a closed source effort. So everything that I presented today is fully disclosed in my papers, there is a source code with MIT license for all the pieces that I just showed, MIT meaning that can be used for commercial, non-commercial usage. The wild meshing efforts for creating tetrahedral and triangle meshes, this is something that many of you might have already used, it’s what is used in anthropology, one of the ways that anthropology uses to export tetrahedral meshes from their designs. We have both a simulator and a set of tools to integrate in your simulator, all our work on collisions, this is something that was mentioned yesterday in the talk from Carbon, that we use for their simulation of a compression of the microstructures.

21:53 And we also have a full-blown non-linear finite element solver that includes contact, friction, and is entirely differentiable, so it can be used for the shape design application that I showed before. All of these come under MIT license, and I will be very excited if anyone in the audience wants to use them, please do reach out, and I will be even more excited if there was a way to get support in form of like development time or some other kind of support to allow me to continue to work on this open source effort. There is a course that is online, it’s on YouTube, it’s I think four or five hours if you want to take a look at this, we are going over a few of the demos that I showed today, showing you how to implement it in Python. All our code is written in C plus plus for efficiency reason, but we wrap it on Python for the purpose of making it easy to to use.

22:48 And so to conclude, the, I think there is way too much knowledge needed to use existing simulation method and design method for the purpose of creating some new artifacts, I don’t think it’s interesting. But for example, if someone cares about how much a beam bends, to know how to mesh it, I think this is just a technicality that should be hidden, and there is no point in making it exposed, as long as as it works. So my hope is that it’s possible to actually get rid of these more difficult way of proceeding, and switch it to something that it’s easier, where you only care about the parameters that you actually need for the specific design problem that you’re working on. And in some sense I think the major success in the machine learning community was pushed forward by the availability of fantastic tools like TensorFlow and PyTorch that allow a high schooler to train a model on his own images, right, it’s so easy to use these tools because a lot of the hard problems are already solved and we can be solved as a community.

23:58 And all this code is available online, so that many people can contribute to it, of course you can build commercial products on top of these two, there are plenty of examples, but the fact that the core technology is completely free, I think really boosted this. And so my hope is that on a different space is possible to solving PDEs, solving problems that are constrained from PDEs, and use them for a lot of exciting applications in in this community. So I would like to thank all the sponsors, and I’ll be very happy to discuss about any of these in remaining of a conference. Thank you very much. Thank you.

More from CDFAM NYC 2023

Design Computation Human

Design Computation Human

Onur Yüce Gün · New Balance

Mud y Robots: Keynote Presentation

Mud y Robots: Keynote Presentation

Ronald Rael · UC Berkeley / Emerging Objects

Roadmap to Replicators

Roadmap to Replicators

Neil Gershenfeld · MIT Center for Bits and Atoms

Scaling & Standardizing Patient Specific Design

Scaling & Standardizing Patient Specific Design

Adam Wentworth · Mayo Clinic

Computational Design Ops

Computational Design Ops

Ronnie Parsons · Mode Lab

Overcoming the Design Dilemma

Overcoming the Design Dilemma

Ryan Carter · Johns Hopkins APL

Generative Design & Digital Manufacturing at NASA Goddard

Generative Design & Digital Manufacturing at NASA Goddard

Ryan McClelland · NASA Goddard Space Flight Center

Digital Materials – Geometry and Data Combined

Digital Materials – Geometry and Data Combined

Marek Moffett · General Lattice

Ten Years Teaching Additive

Ten Years Teaching Additive

John Hart · MIT

10 Years of Applying DFAM

10 Years of Applying DFAM

Tim Simpson · Penn State University

MfAM Walks so DfAM Can Run!

MfAM Walks so DfAM Can Run!

Chelsea Cummings · The Barnes Global Advisors

Informatics and AI for Materials Development

Informatics and AI for Materials Development

Wes John Alder · Citrine Informatics

Predicting Fatigue Life in AM Metal Parts Using DigitalClone

Predicting Fatigue Life in AM Metal Parts Using DigitalClone

Behrooz Jalalahmadi · Sentient Science

Designing Material Properties with Advanced Toolpath Generation

Designing Material Properties with Advanced Toolpath Generation

Alex Roschli · Oak Ridge National Laboratory

Design Automation Using Machine Learning & Computational Design

Design Automation Using Machine Learning & Computational Design

Rhushik Matroja · Cognitive Design Systems

Learning to Generate Shapes

Learning to Generate Shapes

Karl Willis · Autodesk Research

Design for Artificial Intelligence

Design for Artificial Intelligence

Chris McComb · Carnegie Mellon University

Register for Updates and Discounts on CDFAM events.