0:00 Okay, hello and good afternoon, all right, I’m super excited to be here, my name is Elissa, I’m from Metafold 3D, I’m the CEO of Metafold, I’m also a mathematician, although sadly I don’t get to do a lot of mathematics these days, but my amazing team does great mathematics for me. And today I’m going to be talking about the idea of geometric infrastructure for additive manufacturing, and I’ll explain what that is and why we think that’s important at Metap, but first, as a mathematician, I am going to start with a definition, and this is just infrastructure.

0:35 Infrastructure we can define as the underlying foundation or basic framework of a system or organization, and to look at an example, I hope this isn’t too controversial for the the resident Berliners here, but I think there’s probably no better place to talk about infrastructure than here in Germany, and I’m going to use the Berlin transit system as an example of of good infrastructure. So what does the Berlin transit system do, what is its purpose, its role in life, I would argue that its role in life is to connect people to places. And my next question is, what are the requirements on the system for it to do this effectively, it must be robust, meaning it must be reliable, strong, functional, it must be accessible, so people need to be able to use it, know how to use it, we need to be affordable, and finally, my favorite one, it needs to be scalable, it needs to support growth, because cities change, they grow, populations change, so it must be extensible.

1:47 Obviously I didn’t come here to sing the praises of the Ubon, even though I do like it, I want to, I want to talk about geometry, so let’s look at a definition of infrastructure for geometry. I would des describe geometric infrastructure as the mathematical foundations for geometric computing, and I’ll use CAD as an example of this, so, and you might notice also the the similarities here to the definition of a GE geometry modeling kernel, so what is the role in life of of CAD, I would argue this is to connect ideas, sometimes I think about the the geometry of the imagination, or sometimes engineering objectives, we’ve heard a about that today, to connect those things to a physical product or a manufactured good.

2:37 And I would also say that the requirements on geometric infrastructure are very similar to the requirements that we were just talking about at reg regular infrastructure, it must be robust, accessible, scalable, and I hope, and I think we all share this hope, that the additive manufacturing industry is growing, so that scale piece is really important. And so in the the rest of the presentation I will share how we think about these ideas at Metap, and what they lead to, so they, they lead to new kinds of physical parts and products, they lead to new kinds of software applications, and of course new ways of working with 3D data.

3:17 Beginning with the first principle, robustness, and just to level set us here, Metap really offers two products, for those of you who aren’t familiar with it, the first is a web application which I’ll show you in a minute, the second is our API, and I’ll also talk about that later, but both draw from our own cloud-based implicit geometry kernel, this is our geometric infrastructure, right. So this is a very sped up version of our web application, this is happening in the browser, we handle very high geometric complexity, these are all based on implicit modeling, which we’ve seen great examples of here at CDAM, and, you know, just very basically this is the idea of represent, presenting shapes with functions, not with surfaces.

4:05 And we also offer, I hope maybe you caught it at the end, we also offer some mechanical simulation, compression simulation, right in the application, which I’ll I’ll talk about a little bit more. So, but I want to talk about robustness, and to motivate this I’ll look at a couple examples in the biomanufacturing space, so these are two of our customers, one is Salia, there’s a Swiss startup making cell scaffolds for lab grown meat and fish, and the other is SWRI, which is a research institute that is making a bioreactor, which is totally mind-blowing, and that it grows human stem cells, so this has the potential to really transform the regenerative medicine industry, this was recently featured on the Cool Parts Show, if you want to learn more about that, by the way.

4:50 But both cases involve cell growth, and in both cases these structures, you’ll probably immediately see that implicit modeling is the natural choice, there’s so much surface area, so much complexity, and by the way these are very small samples, the goal is to be go much bigger. But here’s the thing, both Salia and SWI need very precise control over the size of the the pores and the channels in these structures, because cells are really small, they’re trying to grow cells, and they need to fine-tune those geometries to optimize their cell growth. So a key challenge for the geometric infrastructure supporting this is that, how do we make implicit modeling very very accurate, or maybe put another way, how do we make 1 mm in an implicit model really mean 1 mm.

5:46 So the the role of an implicit function is to describe the locus of points on the surface of a shape, but it doesn’t actually tell you anything about the other points in the space, and this is essential if we want to give any dimensionality, any thickness to things. So the heart of the problem is really this distinction between a generic implicit function and assign distance field, to give some sense of the kind of what can go wrong when you just begin building up implicit functions by Boolean operations, this image is from a recent paper by Zoe Marner and some colleagues.

6:23 And at the top we have a a true STDF, and at the bottom, this is an implicit function, or, they’re a pseudo STF, and the pseudo STF, by the way, happens when you just kind of naturally boolean together a circle and a square, in in the course of things, and you see that under dilation, things look okay, and under erosion, they, they diverge, they’re different. And in fact the pseudo STF doesn’t give you correct distance values, but this is, is not just about erosion and dilation, this is also just about dimensional accuracy, about thickness, yes.

7:01 So yeah, we at Met Aold, we see implicit modeling kind of like this, it’s all, it’s all fun in games, we’re swinging gyroids around, and then we need to shell it, and we want to do it really precisely, and shelling again is really important, because we live in the real world, the real world has thickness, manufacturing is the ukian metric. So this was one of our real challenges, how do we make the conversion from an arbitrary implicit to assign distance field very fast, very accurate, and and in fact fast enough for interactive CAD, so we have a new algorithm for this, which I won’t go into the details of, but it will take any sampled implicit function and convert it to an SDF fast enough for interactive CAD applications.

7:51 So this geometry on screen is really an extension of the previous 2D geometries, and you see that the erosion is working at as what you expect, so this is, this is good for the dimensional accuracy that our customers need, it’s also good for that kind of expressive shape space that we hope to get out of implicit modeling. Okay, so the second principle of infrastructure I want to talk about is accessibility, RTech is fundamentally cloud-based, and we’ve taken a hard line here, and there’s lots of good reasons, and David yesterday had some great arguments here, but we believe businesses should f focus on maintaining their core IP and their innovation, and not maintaining their compute infrastructure, and also of course there’s a, you know, we can concentrate compute in the data centers and be very efficient.

8:38 This way, of course, being cloud-based also opens the door to all kinds of different downstream recipients of what is effectively a link, so whether it is software, whether it is 3D slices being streamed directly to a printer, whether it is just your colleague who doesn’t want to, you know, get an email with a massive STL file. But, okay, we don’t just put our geometry kernel on the cloud, it doesn’t work that way, geometry on the web is hard, especially editable geometry, how do we make it possible to actually manipulate geometry and do the kind of CAD moves we want to do on the web.

9:20 So our key question here is how fast can we make that request to result loop, and the answer right now, our current best is, this is a little conservative, but on a fast internet connection it’s it’s fast enough for interactive rates. And so this is our idea of geometry streaming, you can think about this as a kind of compression, a kind of extreme example of this is, this is a little preview of our new progressive renderer, so this principle is about making the the details you need available when you need them, the idea of a progressive renderer is that the longer you look at something the more resolved it gets, the better the detail becomes, of course we work on sampled implicits, lower resolutions are faster.

Okay, but we wanted to increase the effective resolution without the implied cubic increase in weight time, so in this case the client device can run rock steady at 60 frames per second while the other rendering compute is happening on the cloud, and all of this, of course, is essential for delivering this sort of smooth and and responsive visual experience in real time, and also just giving people the the validation they need with their eyeballs that the shape is what they expect it to be. The other part about accessibility is making our geometry infrastructure available to other people, to you all, and so we made our web app with our geometry kernel, with our, with our API, but we also make this available to other people building software, including commercially licensed software.

9:34 So this is, this is useful for internal applications, one of our founding observations was that many large companies end up building their own software, or in other words every hardware company becomes a software company, but it’s also useful to commercial applications, and this is part of the infrastructure point, so AM is growing, part of the commercialization of 3D printed parts is being able to get them out into the world, and software is often the mechanism to do that. So as an example, this is a a web application that we spun up as a demo, and please, like, we are not orthodic, so don’t look too carefully at the geometry itself, but the point is this is a little web app that we built with our SDK, the app takes in patient scan data, in this case feet, it prompts the user to select some points on the geometry to guide the geometry generation, so once again doing this on the web is not something that is straightforward to do necessarily.

12:07 It autog generates the geometry using the implicit kernel, and then allows the user to export, in this case, 3mf. Conservative estimate suggests that if you wanted to build this functionality entirely from scratch it could take you months, if not up to a year, you know, just getting geometry on screen, creating it, editing it, rendering, exporting, importing those scan data, all this stuff, with Metafold we spun this up in just a couple of days. So the way to think about this is that Metap is just one of a a collection of web APIs that make it easy and modular to build products, to build applications that support your your workflows and to commercialize your 3D printed innovations, we also work with contract manufacturers who want to connect this directly into their internal job management systems.

13:06 And this leads naturally into the final principle of geometric infrastructure that I want to talk about, which is scale, and again this is really the the kind of goal here is to scale the additive industry, so AM is growing, the need for software to support manufacturing is growing, which we’ve seen so much here, which has been wonderful, and this last application I showed you, know, looks at how you might be able to commercialize a 3D printed innovation. So this, my first example of scales around data, and this is actually a small data example, but we know, and we’ve seen so much here, that optimal structures, finding optimal structures relies on this feedback between data generation and simulation results, this is a informed by high quality data.

13:54 So this is a very small design of experiment example of how a Metaphor user can use our app together with a Python script to quickly generate a little data set and a design of experiment to test a hypothesis. So in this case my colleague was interested in the the effect of different lattice geometries and relative gradations on the energy absorption of the lattice, so she created this, again, small data set, 28 different lates, with four different lattice geometries and seven different grading density profiles. The next step was to run compression simulation, and I’m going to say meshless, I know that’s controversial, but it is, there is no tetrahedral mesh involved, let me at least say that.

14:36 So this is a full quasi static compression test which runs right on those precise exact SDFs, so you don’t need to wait for meshing or introduce any kind of lossy data translation there, so our solver does full compaction, full densification, it handles all the contacts with friction, and we support a wide variety of material models, including hy prolastic materials, we recently incorporated some great materials from BASF, EOS, and Carbon, these simulations were done on the carbon EPU 45, which is often used in the footware applications that I’m seeing on many many feet here.

15:12 And so doing this small experiment allowed us to draw several conclusions, so conclusions aren’t so relevant, the point is about the data generation, but anyway you can see here that in this case geometry had a a significant effect on the specific absorbed energy, and the neither geometry nor the relative density influenced the densification strain. This is a small study, a simple study, but it gives us an example of how we think about the data generation capabilities at scale, and certainly we’ve heard a little bit about people training models to at least interpolate, if not optimize, but this is certainly possible with this technology.

16:01 We can also handle much higher complexity geometries than the example I just showed, again, this is the benefit of using, you know, no tetrahedral meshes, we heard a lot about those challenges this morning. So this is a a beam ified gyroid, just flowing beams over the surface of a gyroid, and you can see that we’re able to simulate this using our our approach, these, what I’ve shown here, happen to be pucks of lates, this is absolutely not not the only thing you can simulate, we will happily simulate any part, and the the magic is really in handling higher complexity geometries, but of course if you’re studying the basic properties of meta materials it’s a great tool.

16:44 Okay, so I am going to go here, I actually have been a little surprised we haven’t heard so much about AI this time, which is kind of cool, as I just said, you can create larger data sets to train your your learning models with the approach which I just outlined, but sometimes I think my position on on 3D and AI is a little like this, it’s, we’re not quite there, it is still hard, it is still hard to generate 3D data using, using AI, and 3D data just is so much more heterogeneous than 2D data. It’s critical, then, when we’re doing these any models that involve these data sets, that the the data is very targeted, of course there is one way where we meaningfully leverage AI, and that is in the the recent LLMs and the ability for LLMs to write code, so this is relevant to to us and to anyone with implicit geometry capabilities, because it is fundamentally code and text based.

17:46 So this is a little recent proof of concept for a kind of design co-pilot, using, using that function, and you can see that we begin in the application, and then we open up this little co-pilot, and we ask it to populate the box with a whole bunch of of donuts, then you need some prompting to kind of rotate those things properly, but we’re going to use those to create a new, a kind of a new meta material, so by booing bullying them together, and then coming back into the web interface we can do some shelling, apply a gradient on the side of those T, so that we get a a a gradient based, or sorry, gradient on the the infill here, and then finally we end with a a simulation.

18:42 Okay, so putting it all together, we’ve looked at these three principles of geometric infrastructure, and how we think about them at Metap, so robust, accessible, and scalable, those are the things that we’ve been thinking about for the past few years. It all begins with our GE kernel for implicit geometry, from there we have geometry creation, geometry modification, the conversion to exact SDFs which give you that dimensionality you need, and our simulation solver, our geometry streaming protocol is what allows us to communicate this efficiently over the web, and our graphics engine handles the the rendering of the geometry, all of this is exposed in the the API, and we build our own web application for design and simulation on top of our API.

19:36 From there, our customers, they make parts, they make new kinds of shapes leading to incredible innovations, they also build data sets, they build applications, and they stream directly to printers, this is all based on the the server client architecture. So I’m super excited to see so many people here at CDFM working on related topics and pushing the boundaries of digital tools for producing physical things, I would love to answer any questions that you have, I would also love for you to try these these tools out, so you can just log on to our app, it is free to try, and similarly the open source SDKs are there for you to use, love your feedback, and thank you so much for your attention.