Okay, great. Hi, my name is Andrew Sink, I’m on the application engineering team at Carbon. I help our customers turn products into mass manufactured products using the Carbon platform, and I’m joined today by, my name is Ruiqi Chen, I’m a software engineer and I work in the intersection of computational geometry, mechanics, manufacturing, graphics, a little bit of everything. So today we’re going to be talking about our road to making 1 million custom parts every month. We’re going to talk about what we learned on the way there, what we’re learning right now, problems we’ve solved, and where we think we’re going to see the future. But just to start, added manufacturing should enable mass customization of any product at scale, using a set of tools that are repeatable across products and industries. Kind of a nice blanket statement, but basically what we’re trying to say is we’re trying to develop pipelines that are portable, workflows that are portable, that we can move across industries.

0:51 So for those who aren’t familiar with Carbon, Carbon owns all three legs of the stool: we make hardware, software, and our own materials. Our hardware is our printer line, so our M3, M3 Max, and L1 3D printers. Software is called Design Engine, which we use to make lest parts, we’ll be talking a lot about Design Engine and customization today. And we also make materials, so those include strain rate sensitive elastomerics, high temperature rigids, and dental specific materials. So what’s on the path to get us millions, millions of plus parts to scale up? Well, it really starts with step zero. We’re going to talk about steps one through four later, but what is step zero? Step zero is, you need an application, you need some market fit, and I love, someone else, you know, earlier presented, like if you don’t have a business case, why bother, right? So where we saw the first business case, especially you know as you guessed it, really much in the dental, dental space. Why dento? Because it really, you know, exemplifies what additive is really good at, which is mass customization.

1:59 So what we set out first is to try to solve a lot of these scaling up problems specifically for the dental space, but having software that we write that’s applicable to not just dental but to other products. So specifically for the dental space, what we’ve done is we’ve developed pipelines such as automated packing that can beat human level packing densities at this point, as well as robust slicing of meshes, you know, people are talking about manifold, water type meshes, we give us junk, we can slice it pretty well. And just last year, in 2023, we processed through this, you know, packing, slicing platform, 360,000 builds, were each build is composed of, you know, maybe 20, 30, 40 denture or thermal forming molds, all packed into one build. This equates to about 10 million parts that were all printed on Carbon systems.

3:00 So in 2024, so this year, we’re now launching what we call the Hollow Model V2 platform, which is not only are we doing a lot of this print prep work for dental models, we’re also doing geometry processing and modification specifically for our Hollow Model workflow. So the Hollow Model workflow is, we’re essentially taking these, you know, thermal forming positive models that, for the most part, take up a lot of resin, they don’t need to be fully solid, we hollow them through an automated, fully automated G Puu pipeline, extremely robust. And not only does it just hollow it, also adds, you know, vent path so we can tailor the manufacturability specifically for Carbon processes, specifically we have to add resin drain path so that it can spin process properly. So up to this year so far, we’ve processed through our Hollow Model V2 platform 760,000 parts, and later this summer we’re aiming to be processing 1 million to 1.2 million parts per month. It’s a lot, you can imagine our aows cost, it’s a lot of money that we are also spending. So there’s a lot of computation, how can we scale this up, you know, this technology, because this, this is a cool industry now, right? We, we are in a case where we’re not just talking about one or 12 parts, we’re really talking about what is the technology that’s needed to get us there, to mass scalability.

So finally, to, you know, finish the mass scalability side of things, it’s not just all about software, I know this is CD F, but it’s not just all about software, it’s all about hardware. With Carbon we’re also focusing on hardware automation, so specifically still looking at the denture automation side of things. We’ve released our new automated printing set of tools called AO Suite, AO for automatic operation, which AO can also stand for always on, because as you get it, as you guessed it, this printer is always on, it prints 247. There’s a kind of a, you can see in the videos there, there’s a knight that automatically slices all the parts off after they print, there’s an automatic washing step, and then it all gets dumped into a bin where a technician can pick up, you know, once in a while he might have an email notification or something. So this product was announced really recently and we’re really excited to start shipping the AO Backpack and the AO Suite of hardware automation.

5:20 Okay, so now we know we have some level of automation in the dental industry. So we take a look at some of the consumer products that car is created. So right now we’re looking at over 4 million lested parts produced, so those are parts that can be purchased at a store. I bought these shoes that I’m wearing at a store, it was a really cool experience, as somebody who’s been involved in additive for a long time, seeing something that was 3D printed that was being sold on the merits of his performance, not necessarily its novelty, was pretty cool. So we have all of these other success stories as well, and so we’re starting to think, okay, so how can we bridge this gap between custom parts and and our our stock product offering. So Richie’s going to talk a little bit about the CD, computational design, part, and then after that I’ll talk about the DAM, and we’ll kind of bridge them together at the end to see how this all comes together.

6:09 So for our performance lattice applications, it really starts with getting both a qualitative as well as a quantitative feel for what type of mechanical performance your application needs. I think the biggest thing I’ve learned as more of a computational engineer is you have to be able to bridge the gap between people who are not necessarily communicating terribly quantitative, like you know, maybe I know the modulus of steel, but what does that number actually mean, what does it mean in your hand, right? So a big part of, you know, design for additive is, at the end of the day we’re humans, we’re going to feel the parts, you need to be able to communicate this. And you know, a big part of answering these questions of, you know, what do I need for in terms of properties, you know, start off with the question of, am I designing for a marshmallow, a pillow, or a hard hat, where are we in the space of mechanical properties. From there you can really get quantitative, you know, pick out your resins, certain resins might do better at damping, others more for energy return, look at your particular lattice is for different, you know, aesthetics or different types of mechanical response, maybe you want something that buckles more so you have a basically a constant stress profile, or maybe you want something that’s just a little bit more nonlinear, like a foam response.

7:18 So with, you know, this in mind, we developed a tool called Lattice Search, which really lce search focuses more on the quantitative aspects. But what we done is developed the lattice library, and the lattice library is just composed of a bunch of lates with different parameters, all the parameters are continuous, including the cell type, but cell type, cell size, strut diameters, these are all continuous parameters that you can, you know, play around with. This infinite space, and of course we can’t simulate an infinite space of lates, so we substan it, we simulate it through, I’m going to, I’m going to say this in order, it is an implicit nonlinear dynamic simulation that can handle postbuckling as well as hyper elastic materials, so not just your linear elastic materials, as well as self-c contact. So all in all it’s solving a pretty nasty computational problem, and we’re proud of that, but I will say, like, this is tough, and I’m really excited to hear some conversations later, later today and tomorrow about, you know, advances in the computational space, because I’ll be honest, like these are really, really tough problems for computation mechanics.

8:23 So where we found some of the best use in computation mechanics is really focusing on building these, you know, building blocks, getting these building blocks of mechanical properties which we can simulate in a reasonable amount of time. From there, using things like experimental validation as well as machine learning, you can start filling in the gaps of, you know, interpolating between, as you tune these continuous parameters, where do you get in terms of your mechanical property space. Okay, great, so now we have a big library of lates, we have, it’s searchable, it’s indexed, we’re ready, right? And so this is where we have first contact with a customer. Somebody comes to us and says, “I have an idea for a project,” and the email they send us, they say, and this is a pretty good example of what, like, 90% of incoming emails look like, it’s somebody squishing a block of foam saying, “I’d love it to feel like this,” and then we get a 2D picture of a l structure, “but I want it to look like that.” And that’s the baseline. And so now we have all of this data that, historically, this was a pretty, pretty, yeah, gnarly challenge, this is tough. So now we have all of this data that’s indexed, so we’re able to reduce the number of digital iterations that we work through.

9:27 So when we think of iterating on, on any consumer product, there’s digital and physical iterations. Digital iterations for us, they come with a bit of a catch, the material has to perform very specifically, we’re typically replacing a piece of foam that has very specific characteristics that may not be documented, and so we’re trying to understand, will our lattice perform. Once we have the digital iteration finished, we move on to physic iterating, and this is where we start to think about printability. So on the application engineering team, I think about printability a lot, and this is where we start to get those last minute requests, like, can I add some venol, can I add a logo, can I add a texture, things that aren’t necessarily part of that initial lattice. So we start to go through all of these physical iterations, what are we trying to solve for effectively, and ultimately, at Carbon, we’re experts in additive manufacturing, and we expect our customers to be experts in their domain, so we look to them to tell us what defines success, and that’s the outcome that we’re trying to drive towards.

10:23 So once we have a part that’s ready to go, and it prints well, and it looks good, in this case it’s like a nice breezy back rest for a backpack, we ready to take it to manufacturing. And this is typically, if you haven’t printed at large volumes before, this is where the journey typically stops. On the CAD side you get to something that’s printable and you think, okay, we’re done, right? So not quite, this is where we start to think about manufacturability testing, internally we call this stress testing, and this is something that’s really exciting to me because it represents this very, very small niche in additive manufacturing that isn’t exam very thoroughly. So what you’re looking at here, you’ll notice that there is a strut lattice with a red outline around it, in the CAD you’ll see that nominally it printed okay, so what you’re looking at is a very thin beam that’s supporting much larger structure, and it looks like it printed okay, but then under stress test it failed.

11:13 And so stress testing is where we’re intentionally lowering the exposure of the UV light that’s exposing to the part, to make it weaker in the green state. And the reason that we do this is we want to test to discover the difference between a print related failure and a design related failure, because print related failures, one-off problems, they’re really easy to solve for, you just turn the printer off, turn it back on, the problem takes care of itself. Design related defects, as it turns out, they’re a little harder, because this little piece might work eight out of 10 times, but on printer number 10, that a certain operator doesn’t necessarily clean fully, maybe the resin, there’s a little bit of lot to lot variation, it’ll fail consistently. So you’re going to throw a lot of flags, it becomes a real challenge. And so internally, what we do is, is once we have it designed that we feel confident about, we put it through this manufacturability testing. So I know it’s easy to print one thing in a lab, when you’re printing 50,000 of them at a facility you want to make sure the first one and the last one are similar, and that’s really hard with a last americ materials.

12:10 I can talk about this slide way longer, we’re going to keep moving, but if you have any questions about this, like, let’s talk about this later, yeah, go talk to us, we have a bunch of, we have a booth in the back with a bunch of parts you see here. But really, like, you know, I love the, of this conference, it’s CDFAM, right, because really, condu, we’re connecting CD, computational design, a lot of things done in a vacuum, with manufacturability, where now you have things that just kind of don’t go right sometimes, and physics can’t really explain it, right? So this is where there really needs to be a balance between computational tools and essentially iteration, iteration has to be part of the process, we just haven’t been there, 100% where we can do everything digitally yet, we’re not there yet, maybe one day, maybe one day.

So now you’ve seen, at least, a step through what it takes to get all the way through the process, from conception to manufacturability stress test, so that we can actually start scaling up in our factories. So what does this process look like in terms of what a customer sees? So I will say this, C, Carbon has a complex business model in that we have all sorts of customers with different requirements, different types of input formats, or just, what are they trying to get out of this, right? Are they trying to make millions of dentures per month, or are they trying to, you know, make 1,000, you know, custom limited edition shoes or something. So lots of different types of customers, and we want to develop a platform that’s able to support that.

So what we’ve done is develop software where we try to work and meet the customer as much as possible in the middle. We start with a Carbon DLS API, which is flexible for customers that want to really integrate directly with our software, alternatively, we can also, customers can go through our platform, through Design Engine, which is our web UI for doing latticing and design. So going a little bit deeper into what makes up the Carbon DLS API, it really all starts off with the concept of operations, which is just, you know, very similar to like Blender nodes, or these nodes in Grasshopper, right? So these operations design Define something you want to do in computational geometry, it might be, you know, latticing, adding a textures, or it might be print prep, like adding engraving, slicing, doing the actual layout and packing, or especially for us I would do want to mention, like, flattening, like, because our printers work best when parts are actually flat, and it also lowers costs, flatting is a huge, huge, huge tool for Carbon DLS application.

So and this is just a subset, we actually have, experimentally we have like 90 operations, and a lot of them are like, you know, like you know, like they work 80% of the time, 20% they’re, you know, some hiccups, but we’re working to improve that more and more every day. So it starts off with operations, and what happens when you start, you know, cascading a lot of these operations together, you get a, what we call a model program. So a model program is just, you know, it’s a directed e cyclic graph, this is a pretty simple direct e cylic graph, but it’s just the sequence of operations that Define, you know, some sort of workflow that you want to do, right? So this would be a theoretical workflow for designing, let’s say, an inso application, where you start off with data, which might be a design space, saying the envelope of apart, along with some pressure map data from, let’s say, a specialized foot store doing a foot scan.

15:29 From there it goes into the Carbon DLS API, it goes through a latticing process to create the lattice, add some feature edges for, you know, aesthetic purposes, turn it all into a triangle mesh, sorry, Duann, we use STLs, I’m sorry, but this all gets then connected down to, you know, maybe some additional features like adding patches for texturing applications, and then doing a lot of the print prep work I alluded to earlier, actually oriented packing, it on our platforms to get a printable build, and then this gets sent to automation, and it gets printed either by us, by our factories, by customers with our printers, or through the Carbon Design Network, which allows customers who don’t want to necessarily own printers to also be able to access our technology. So a model program is just a a sequence of operations in a directly cylic graph.

16:23 In our platform, these model programs then get sent through our software architecture, which is all microservices based. So maybe a high level overview, the software architecture, once it goes into our system, it goes into a server that really dissects these model, ser, model programs, we call that Model Program Service, it dissects these into individual operations, and then estimates, you know, roughly how much time each operation needs, how much memory it needs, any specific hardware such as a GPU or whatever, and then that gets sent to a worker load balancer, that then gets sends these individual operations to our AWS clusters, where the jobs are actually run. And then we store everything in a model database and cach all of the results, as well as the inputs to the results, such that we never run the same op, we never run the same operation more than once, as long as the inputs are exactly the same.

17:17 So now we have this fairly robust pipeline where we understand the software architecture, we know that we can make custom parts by changing the inputs, and it’s we have these reliable systems in place, as well as printability tests, to make sure that parts are validated. And so we start to put together the pieces for how to make custom parts at scale. And so what do the next generation of custom parts made with added manufacturing look like? And so we need all of these pieces to kind of put this puzzle together, the materials, the software, the manufacturability testing. And so when you get all of these pieces and you put them together, you get a result like this, where two out of the top 10, I’m sorry, four out of the top 10, and two out of the top five helmets in the NFL, rated by the NFL Players Association for safety, and these two at the top, the M, the VIIs Matrix ID, and the Rell Precision diamonds, these are both custom helmets, so these are made with that same custom pipeline.

18:09 So we know that this is a system that works, and we know that we’re doing something correctly, because we have all these pieces to put together. So from here the next step is, okay, what are all of these other industries, footwear, helmets, boxing gloves, bike saddles, things like like that, where we have the opportunity to to grow. And so that’s kind of where we’re focused right now, that’s where we’re thinking the future is for additive manufacturing, we see a lot of success historically there, and we also see a bright future. So with that in mind, thank you very much for having us. Yeah, oh.