CDFAM NYC 2023 · New York · 14–15 June 2023
Reinventing the Wheel – Representing & Manufacturing Geometry with a Program
Abstract
With a rich background in research and a focus on geometric representation, Mark explores the intricate balance between design freedom and computational complexity in the manufacturing industry. He delves into the representation and analysis of complex parts, such as lattices, and highlights the efficiency and advantages of the 3MF file format.
The talk also covers innovative applications of new mathematics, digital twin technologies, generative design, and the potential of code in representing geometry. Mark shares real-world examples, discusses the challenges and solutions in 3D printing, and offers a glimpse into the future of design and manufacturing.
Whether you’re a designer, engineer, or enthusiast in the field of additive manufacturing, this presentation, recorded at the renowned CDFAM Symposium, offers valuable insights and cutting-edge knowledge.
Transcript
From YouTube’s automatic captions, lightly cleaned; expect some errors. Each timestamp opens the video at that moment.
Read the full transcript · 4,937 words
Foreign. Just a quick video, I’m not going to run this all the way through, but just to show you what we’re, we’re kind of up to here. I can get rid of the contact us. Some fashion is particularly exciting right now, an additive. A lot of vendors like to say you get complexity for free, that’s not as true on the design side. One of the common problems for any design engineer in manufacturing industry is to strike this balance between the design freedom and the computational complexity. Our invention offers a lot of nice advantages. Right now there’s so many things that designers and engineers cannot create, or it’s just too difficult to create with the existing tools. So we’re giving them new representations, new methods, new, new functionality that will allow them to create these, these new amazing products.
1:18 But not such problems is the design representation and analysis of complex parts, and for example, parts that may have lattices in their structure. Think of Eiffel Tower, and that is one example of a larger structure at a macro scale. So the Eiffel Tower consists of arrangement of beams that are connected to each other in a periodic fashion. In fact, 1B may have several other smaller beams, and in fact that is the case with the Eiffel Tower as well. Now imagine increasing the number of beams from a few thousands to a few trillions and connecting them in many, many more complex ways. So in terms of lattice microstructures for 3D printed part, that is the level of complexity that we are dealing with. So previously these, these designs would take, you could take hours even to load them into the into your design software, and with our representation we can do that in seconds. We’re bringing that down to kilobytes in sizes, so it’s a ridiculously huge savings in representation size. We could have looking lattices, but all right, so that’s enough of the video. Here you didn’t come to see videos, so let me put a presentation up here.
2:54 All right, so I’m, I’m Mark Burhoff, I work for, for Siemens Technology, which is basically the research group of Siemens. So I’m really, you know, I’m really excited for this group here to be here, like doing was saying, it’s, there’s so many people that I’ve been wanting to meet that I finally got to meet, there’s so many people I met the last, the last day. It’s just wonderful, it’s really cool. Usually I would ask, you know, since we’re in New York City, who went out and went to see Times Square or whatever, but this group’s a little bit different. So who, who went home last night and did an interface to Chat GPT for your software? So I, I had that idea, I thought, I thought it was, I thought it was really smart talking about that a month ago, and I learned yesterday half of you all have integrations for your products already. So I’m, I’m a little bit humbled now, I’ll have a lot more to, to work on here, I have to raise my game a little bit.
3:55 So, so if you don’t know Siemens, if you came in on a train today, or, or leaving on a train today, it was probably pulled by a Siemens engine. If you work in the medical field, or you’ve been in the hospital or visit your doctor, you probably saw Siemens on a lot of the equipment there. If you design products or do simulation, maybe use NX or Simcenter or some of our, our software there, those are all different business units within Siemens. I don’t work for any of those business units, I, I work for Siemens Technology, which is basically the, the research group for Siemens. So this is other group here for the United States, we’re based in Princeton, New Jersey, so it’s a really easy trip there, and what we tend to do is, we don’t have, we don’t sell anything. The real reason I’m here is willing to partner with many of you, we partner a lot with universities, with other, other folks. We work on problems, you know, for our business units, we work on problems for the government, you’ll see some work we did with DARPA here coming up, and many of you actually participate in some of this, this darker work as well, and, and a lot of it, you know, or we’ll look on some things that are maybe important across all the business units.
5:03 So you’ll see us talk a lot about the industrial metaverse now, not my talk today, but, but with me, again, these are, these are some areas that we work on. I’m more on the, I guess, on the green side here on the, on the left side, but next generation algorithms, which we’ll talk about, digital twin technologies, all these things with generative design, a lot of machine learning. So a lot of our folks are working on, you know, Chat GPT or a lot of the machine learning type things. We do a lot with materials, so we, Siemens owns a lot of added manufacturing machines as well already, so we, we do a lot with developing our own materials, machine technologies, we, we use all of them and, and we love software and hardware. So we have a lot of, you know, clusters, we do a lot with GPUs and other things there. So if you’re looking for, if you’re a researcher, want to partner with us, that’s great, if you’re a PhD student, please come talk to us, we have lots of openings there.
6:05 So this is, this is the, you know, you know, reventing the wheel. So this is a concept that we had a few years ago, basically in talking to, to NASA on the spring tire that they’re building, and this really came about from a few years ago in 2006, 2016. Dark, you know, those of you that are, you know, in your 30s and 40s, remember the biggest debate in CAD back then was whether you want a history tree or not in your Cat system. It was kind of settled down, it was kind of not really interesting, and that additive manufacturing comes around and gets a little more serious, and all of a sudden you could create all these wonderful things that you couldn’t design, and so that, that really made things interesting.
6:51 So, so DARPA, which is, you know, very forward thinking, said hey, maybe we need to fund industry to advance the state of the art in, in design software, how can you now design these, these super complex parts that, that you can’t design today? And of course lattices, metamaterials, all these things kind of come into, into play there. All of a sudden you’re not filling your parts with a header, or homogeneous material, it’s, it could be anything, right? It could be different materials, it could be all these different shapes and so on. So we partnered, in this case we partnered with Palo Alto Research Corporation, Georgia Tech University, Michigan State University, and we came up with some new ideas on how to, you know, it was really three things, I’m only going to talk about one today. It was, how do you represent these really ultra complex project products, how do you do simulation? Now, so you got this awesome complex thing, how can you do a simulation there, you can’t just put a tet mesh on it. And how can you then manufacture these, these ultra complex parts?
7:58 So I have parts now that, you know, if I slice them, well, first I can’t slice them because they’re too big or too complicated, and then I can’t send it to the machine because the machine doesn’t have enough memory for the, for all the G-code that goes there. So there’s still, you know, it’s gotten a lot better since then, a lot of these DARPA folks were PhD students, they’ve gone out into industry, they work with a lot of the companies, some of you may have even worked on some of this. So it’s, it’s really kind of interesting. So I, I’m focused on geometric, geometry representation, so the additive manufacturing people are my best friends right now because they’ve created so many really interesting problems.
8:36 So the, this blue side here, don’t look too closely, I just threw some words up here on things that were, you know, that I remember, but this blue side here was pretty well understood for a while. The new things here that, you know, that’s going on are, are really a lot of fun. So these lattices that we’re doing, 3MF is changing like crazy, we’re talking about metamaterials now, we’re doing things with fields in applied fields, we’ve got implicits, so new mathematics, voxels are suddenly, you know, really in vogue right now, much more than just Minecraft. Topology optimization, so, you know, structural is there, CFD is, is crazy, I put G-Co down here, I, I want to fix G-code, G-code is my, you know, after, after we kill STL, I’m going to kill G-code, but it’s, it’s, you know, still, still growing strong. I put code off here in a second, this is what we’ll talk about. It’s a fun area right now because we, we still have a lot of things that we can’t, that we, there’s a lot of things you could probably print right now if you sit there with the G-code, and write, a really fun area to be working on.
9:51 This, this is actually an old slide, but it’s, it’s, it’s, you know, this, this is our justification for a lot of the work we did, and now it’s, it’s more of a state of the industry type slide. You know, shoes, people are doing shoes all over the place, you know, consumer products, lightweighting is, is a big deal, energy absorption for, for sports equipment is big, woven materials, I think we, we can still do some more things there, but you’re starting to see again where they’re, we’re running, you know, filament through fibers, through, through tweeting prints as well, medical devices, and so on. So really a lot of interesting things going on that, where this is this type of research is important.
10:33 So, who, if you follow me on LinkedIn, I put this post out here, so don’t, don’t answer the question, but for the rest of you, how, here’s, here’s a really simple box, and it’s got why film lines going through the cylindrical wires going left and right, up and down, forward and backwards. How much space would that take to represent in whatever format you like? Any guesses? Well, it’s, it’s huge. If you, if you go back to your, your, your dad’s CAD system, which was only six years ago, if you try to do this with B reps, solid models, it’s huge, it’s, it’s over a gigabyte, and actually I, I get kind of tired of waiting for it, I, I knew I had a smaller model that was over gigabyte, so I know it’s over a gigabyte, but it’s a lot of space. And it’s, it’s really this is why, you know, you see a lot of people looking up new ways to do things.
11:26 STL, 237, so, you know, it’s a faceted model, a little bit smaller, and, and these dials on the side, those kind of represent accuracy. So your B rep is pretty accurate for the STL, right, to get it that small you have to, you lose a few things. Hybrid CAD, and in hybrid CAD, I, I don’t want to call it hybrid CAD anymore because it’s really kind of today’s CAD, and most CAD systems are not looking at B raps as they’re only representation. Most of them are looking at, you know, facet models, they look at graphs, they look at a lot of other things, and in many places they’re becoming equal peers. So you have to work with multiple representation forms as equal peers, you have to be able to take, you know, facet model with the B-rep model and do a Boolean, and subtract things, and have faces that are, you know, matches, and other faces that are, that are precise geometry, and maybe other faces that are, that are some implicit definition. So it’s really you see cab moving that direction a lot.
12:28 3MF is, is awesome. So don, thank you very much. This, this is using the lattice extension there, so what this does in 3MF is lets me represent lattices as a graph rather than as a faceted model, and it brings it down a huge amount. And then we have get text messages, we have our representation here, which is represented as by program, which is tiny, tiny, tiny, and you know, it doesn’t apply to everything, but, but I’ll show you some really interesting things here. And, and really one of the things I want to do is get you all thinking about this too, because I know there’s some really smart people in this room, apparently a lot smarter than me, because I hadn’t thought, I hadn’t integrated Chat GPT in my software yet.
18:21 But you know, think, think about some of this here. If anyone, if anyone can pronounce this guy’s name, that would be great, Call Mo Gorov complexity. I love the, I love the, I love what it is, I, I never mastered saying it yet. This is kind of the concept here, oh, to a certain extent, if you have, if you have this, I guess I’ve always get ahead. If you have a, if you have a string of, of characters here, so here A B A B A B A B, and another one just kind of random characters, they’re both the same length. The, the idea is, what’s the smallest program you could write to reproduce that, that string of characters, at least in this case? And the easiest one is you simply write a function that returns that string, right? But if there’s some pattern, some information there that, that you can use, you can really take advantage of it by writing a function that, in this case, it’s A B, it’s A B.
And so if you’re doing random things in your added machine, this may not work so well, but most of you are not, most of you have some, some, some lattices, especially are very much patterning. There’s patterning, and maybe morphine, which is, it’s almost the same thing as what we’re doing here, we pattern and we morph. So really, you know, let’s, let’s apply this not so much to compressing data, but let’s talk about how we can represent geometry. So this is, you know, one of our examples, and this, this is actually that probably took this about four or five years ago now, this is a, a 3D print that we did, it’s a woven, woven lattice, and the entire representation here is, is on the screen.
19:17 So if you take a picture of that, you ran our software, it would reproduce this piece of geometry. Compare that to an STL file, if I already files to print, the STL file, it would be, be a big box of paper. Here, it’s not really good, if I could did a graph of this, it would still be pretty big. So this is, this is really, really pretty cool. So I, I may be new with Chat GBT, but I have Dolly down pretty good, one of the problems I had at Dolly said, okay, show me the difference between DNA and elephants, and, and this is kind of something I got. So the, this, this program representation is a lot like DNA, right? The DNA is the program for how to build an elephant.
The problem is, you know, lots of, you know, there’s lots of ways, lots of people written code to generate geometry, right? There’s even the CAD systems out there, all they do is you, you write a script and it outputs geometry at the end. That’s not the solution, it’s, it’s, it, yeah, you get a small representation at the end of the day, after you run the program, you’ve still got this giant representation of things, so you still get an elephant that you, you need to somehow work with. And so that, that’s not by itself is not the solution, we need to think about how we can, you know, not, you know, instantiate an elephant every time we want to do something with it.
So this is where they, where it got a little more interesting, we spent a lot more time on the mathematics, there’s a lot of papers out there, so you know, I’m giving you the high points here, but you know, go look at the papers if you’re really interested. If you, if you do a, a CAD system, it doesn’t matter what cat system is, what design system it is, you do it, you end up doing queries, and maybe it’s not called queries, but it’s essentially queries. You’d say, okay, I got a point that XYZ location, is it inside, outside, or on the surface of my geometry, and if it’s outside, what’s the distance from it from the geometry? Other questions, you know, 3D printing, you know, I’m on a 2D cross-section of my geometry because I’m gonna, I’m gonna print layers, right? So give me the, give me the 2D cross section of my geometry, or you know, what’s the weight of my part, how much, what’s, what’s the mass? For NASA, what’s the mass instead of the weight?
Traditional methods, you know, again, we talked about it, it could take hours. There’s one part here that we have, which is a one of the DARPA challenges we had, and it’s kind of an optimized lad, as you can see, kind of the fields running through it a little bit. So I, I created that, and then, then I wrote some code to bring it into our CAD system to load it, and it takes hours, it’s, it’s just so huge and big, and you can easily imagine gigabytes and terabytes of memory for really big parts, especially as these lattices get really small, or if you get really big structures, the graphics get slow. So I know a lot of you are, are struggling with that too, but you know, I got so much data and you need to find ways to, to bring the graphics down to something that’s manageable, so the user is not sitting there with the screen jerking once a second as you try to rotate, but after you get everything loaded, it’s pretty fast, if you’re assuming you can load everything.
What we did here was really an idea of kind of a lazy and local method. So we, we have seconds to load a file, so this big, this big thing is a few seconds, you know, it’s kilobytes of memory needed, which some of that you, you saw earlier, it’s, it’s fast graphics, you know, regardless of the size, and, and it’s a little penalty though, we have slightly slower queries, and, and the reason that is, is because we don’t understand, do it in a kind of a lazy and local way. So lazy means I need some geometry I want to see, so we’ll instantiate the geometry then, but computers are getting so fast, it, it, it, it works pretty well. We also do it locally, so if you want to, you know, look at a, a slice of your part, we just need to instantiate enough of the geometry around that slice to do the slice, we don’t need the whole darn part loaded into memory to do that. Now eventually it will, but it, it just works really well.
19:36 And, and then the other part, you know, some things you maybe need the whole model, right, you wanted to figure out what the mass is of the part, you, you can’t do that, you know, just with, with local things, but it turns out you can, because you can now distribute it across all these computers, and each, each computer, or each EPU, or each GPU is calculated in a different piece of the model. So we, we have some scalability now that we can get the mass relatively quickly by, by parallelizing a lot of these things, and paralyzing is has been a challenge, a lot of, you know, in geometry it’s, it’s relatively easy to paralyze your FEA solvers or your renders or other things like that, paralyzing around geometry is, is a harder, harder thing to do, but this is, this is at least one way that we’ve been able to do it.
20:32 So again, it’s, it’s really geometry representative code, and this is one of our examples here, just I, I backed up. So this, this the nice thing here is this, this part is pretty complicated, but it’s kind of the limit of say B reps, if you want to do something like that, you can pick some other methods that go a little bit farther. We, we got to you saw the video yesterday, Powers of 10, I was watching it hoping it would get to 10 to the 15th, but it got to like 10 to the fifth. We got the 10 to the 15th number of elements in there, which is astronomical number of, so that would be 10 to the 15 beams, let’s say 10 beams per unit cell, so we get 10 to the 14th unit cells. So if anyone compete that, you know, I’ll, I’ll, I would like to follow you and learn how you did it because it’s just amazing in, in being able to do something like that.
21:30 So it’s nice too, because the added guys kept, you know, hey, you can’t hold our geometry, well now I got this, it’s like, can you do it? It’s like, well, you can’t do the resolution I need because I have a lot more detail now on the, on the design side, so I’m ahead of them at least on this, this area here, you can kid my, my additive friends a little bit more.
21:51 So it’s real simple, this, this is a simple example, I, I put this here just to kind of give you a flavor for the idea, it’s kind of trivial how it works, because I think just with this is just a kind of a translation model, so you’re just patterning with exact patterns, that’s a really easy one, but you, we can do rotation, scaling, all these other things too that make a little more interesting. But if you, you, we have this idea of a kernel, it’s kind of like a unit cell, but not quite the unit cell, it’s a little bit different, but what you have is some nodes in your, your kernel, and you define connections, and so you say, this, this node in N1 connects to N1 and the next unit cell, if you will, we’ll stick to that term, in the eye direction, and this one connects M1 in the J direction, this one kicks to N1 in the K direction. So that’s what you see here, so you have a node here at this, this spot here, and it connects to here, here, and here, and then you can, you can start to add other transformations. So here it is with, you know, we’re showing the range in the, in directions in this little model here, so here we have two, two directions, or two in that direction, three, and if you add more and more and more, you kind of see it, it scales out. So that’s kind of the basics of it, and that, that’s pretty easy to understand, right, lots of CAD systems do patterning and other things like that, but I guess it’s a nice easy representation, we can do more things a little more interesting.
23:26 So you saw at the beginning, we talked about the Eiffel Tower in the video, so doing lattices of lattices is, is really kind of cool, or lattices and lattices and lattices, were the, where the lattice beams are actually made up of more lattices. You know, that, that’s really kind of pushing the boundaries right now, but again, we concise representation, you know, you go to any additive manufacturing conference right now and you see this table will have a bunch of lattices to show that they can print lattices, and this table have a bunch of gyroids to show they could print gyroids. So here are gyroids made of lattices, which, so you know, we can do some some really interesting things here, you know, some of these things we could probably do gyroids made of lattices made of gyroids if we really wanted to.
24:13 In a lot of this, you know, I’m talking about lattices, it’s, it’s, it’s the, the piece we kind of push through, and really beam lattices, but this, this works with other things too. We did some, some rockets that we were designing with this representing code, and the ultimate output was, was voxels, but the, but the rocket itself, the definition was a piece of code that did everything. And so, like I said, it’s not just lattices, you know, it’s, it’s, it’s any kind of repeating geometry, maybe it’s not even connected. So here’s some, some Lego pieces, that’s really a bunch of inter, interconnected parts, this one’s fun for my, my grandkids. How am I doing on time? Does that mean I have plenty of time? I get zero. Oh, I didn’t see you, I didn’t see. Okay, so I’m gonna wrap it up, this last slide, I’m sorry.
Okay, good and bad, so let’s last thoughts, you know, program representing with, with what you want, encapsulating equations is great, implicits, you can encapsulate implicits in, in a program really easily, you can calculate, you know, voxels in there, you can get derivatives if you want to do some more advanced things, you encapsulate G-code, this, this is a good one. So we, we if you have an Ultimaker three or above, there’s actually a really nice Linux machine inside that you don’t know that’s in there, so we hack the Linux machine, and now, now our Ultimaker 3 takes a piece of python code and that, and it runs it, and, and then from that python code it lays lily and locally generates a G-code, so it’s manufacturing as well. So it cuts down, you’re not giving this huge pile of G-codes or something, you can give it an algorithm and it calculates, it’s smart geometry, you know, it, it adapts to the machine you’re on if you need to. So let’s say you want to output voxels, and on this machine the voxel resolution is here, but on this machine the box will resolution is a little bit higher, you don’t have to give it different voxel files, you can give it the same program and it will return the voxels for the resolution you need. So all the power of code, which is awesome.
26:24 There’s some bad things, but I’m out of time, so no, no, I mentioned that the bad, bad things too, endless loops, right, you don’t want to put a machine, you don’t want to put your 3MF file in a machine and have it lock up the machine, right, so you have to deal with things, all the problems with code. Bad actors, that this is the one that, this is, I hate, I hate these folks, because I now that you need to have security in there, so you don’t want to put in a file in your 3D printer and suddenly your 3D printer spline and everything you do, that wouldn’t be so good. Lots of solutions, so these are things you know, if you’re on Discord, do on next, next year I want people Discord IDs on their badge too, we’re a lot talking about other things, so there’s sandboxing and there’s WebAssembly and other, other things that are really, really good code reps.
27:11 I, I, I meant to give a shout out to for the, the end top folks there that, that we’re talking about this, you know, that a lot of this came from some common origins with the DARPA project, so the code wraps, maybe it’s an interesting one as well. And finally, I’ll get off the stage. So contact, if you want to connect with me, like I said, we’re always interested in, in working with you all, in, in supporting the community, in, in doing different things. Again, I’m on the technology side, so I’m not really, you know, I can be friends with everybody, right? So Siemens is hard sometimes, because one company will be a competitor, one business unit, but another business unit it’s a partner, and other business units the customer. You know, in my case, I get to be like I said, friends with everybody. So, so thank you, thank you very much.
More from CDFAM NYC 2023

Generative Design & Digital Manufacturing at NASA Goddard
Ryan McClelland · NASA Goddard Space Flight Center

Generate-to-Validate: Optimal Designs and the Need to Deliver Under Uncertainty
André Wilmes · Rafinex

Defining Performance not Just Geometry with Software Process + Material Science
Hardik Kabaria · Carbon

Interspecies Forms: Robotic Feedback Systems Mycelia Growth & Computational Form
Natalie Alima · Adidas / Biolab Studio

Designing Material Properties with Advanced Toolpath Generation
Alex Roschli · Oak Ridge National Laboratory

Design Automation Using Machine Learning & Computational Design
Rhushik Matroja · Cognitive Design Systems























