CDFAM NYC 2023 · New York · 14–15 June 2023
Defining Performance not Just Geometry with Software Process + Material Science
Abstract
This recording is a part of the CDFAM Computational Design (+DfAM) Symposium and features Hardik Kabaria Head of Software Engineering at Carbon.
Presentation Title: Defining Performance not Just Geometry with Software, Process + Material Science
HardikKabaria Head of Software Engineering at Carbon, presents an innovative perspective on the shift from defining just geometry to defining performance using software, process, and material science. His discussion highlights the pivotal role that these three elements play in enhancing the performance and efficiency of products in the realm of computational design and additive manufacturing. Drawing on his extensive experience at Carbon, Hardik shares practical examples and insights into the future of performance definition in additive manufacturing.
This session is of great value to professionals, researchers, and academics in the fields of computational design, additive manufacturing, material science, and software engineering. Those interested in understanding the evolving dynamics of performance definition in additive manufacturing will find this presentation highly informative.
Transcript
From YouTube’s automatic captions, lightly cleaned; expect some errors. Each timestamp opens the video at that moment.
Read the full transcript · 2,698 words
0:05 My name is Hardik. From Carbon. Carbon is a 3D printing company. Today I’m going to talk about how we can we enable our customers to design lattice parts. And there are several examples. I’m happy to show all of that. For the materials and aesthetic because there are as we have seen through the day, designing is not a one variable optimization. You care about performance. You care about the cost.
0:30 You care about which material you want to use. You care about aesthetics. And when we started with 3D printing we the company innovated a new process. And that unlocked new type of material that can be printed at a much higher speed. But it’s almost like we had a hammer and we were looking for a tool. That’s I think about most of the 3D printing industry. We have an amazing technology.
0:56 We are looking for applications. And realization for Carbon was also pretty obvious that unless we figure out a way to make it easy for mechanical engineers who are not on the cool side of the chasm so they can explore things and adopt in the regular workflow, it’s going to be a while before we’ll find more applications. And I believe I’ve been working with for Carbon since 2015.
1:21 Didn’t say that. In my current role I lead software engineering. I still think we are far from it. So we have a long way to go. So I’m very happy this conference is here and there is a realization that we have to still make it way, way, way easier to adopt our cool technology. Where are we? So in that evolution we realized that we cannot just focus on the printers.
1:42 Our first customers was Adidas. I’m wearing the shoes. For those who haven’t seen Adidas is really pioneered this. They were our first customer and very first design of the shoe was made in SolidWorks if you can believe it. And that was a realization we cannot continue this way. And Adidas is still a pretty big customer. They have the chops, they have the engineering team, they have the design team to explore all of these things.
2:07 But we realized that we also want to serve the customers that are not at that size and scale, and we have to help them. We created a design engine, which was basically tool that was used by our internal engineers, our internal application engineers, and we realized that hey, it’s useful. We released it to everybody 2021, and since then it’s been used outside. There are several award-winning products that are developed by our customers that have been launched based on that.
2:36 Couple of different helmets, with military helmets, bike saddles, gloves, backpacks. But if you zoom out from it, the place where we find the really the sort of when it clicks, it clicks because one of the few things that I’m listing here helps. We are better at energy absorption or some sort of a directional energy recovery. We are doing some comfort or a stress distribution that cannot be achieved with, you know, five or six different materials.
3:02 You have to really do some sort of smooth transition. You’re doing mass customization. You care about more than one thing, so it’s like breathability at the same time of stress distribution. Those are the things generally work out in saddles and insoles. And assembly cost is going down because instead of a bike saddle being made with six or seven different forms, now it’s a single part. So those are the key aspects that generally we have seen has to click and it has to be more than one of them for it to be adopted because additive manufacturing still has a uphill battle because it’s a new technology.
3:34 Most engineers don’t know how to use it. And yeah, it’s costlier than most of the other manufacturing methods out there. So one of the things we realized is to achieve some of these things, we had to provide multi-variable or a multi-stiffness part. So that is the first problem we started working on. Many products have been developed on it, but we have rarely talked about it. Partially because we don’t like to go to technical conference, but it’s all right.
4:03 This would be technical enough. So today my talk is not at all about CellC, but what is the under the hood? One of the things here, each of these structures are very different. They will all same bulk material will produce drastically different mechanical properties. And what this picture shows is you can do smooth transition completely controlled by the user or algorithm, what have you. Go from one to another.
4:23 In this case, the topology’s changing, and there will not be any hanging nodes. It will be all compact graph. So that means from the user perspective, they can create a saddle, and they can have three different zones, and they can create a lattice, and they can have different lattice topologies in it. It will all work out. Here’s the math behind it. Like I said, there is we just haven’t spoken about it, but it doesn’t mean it’s secret.
So triply periodic minimal surfaces are quite commonly used in the lattice world. Think about this is what we have done for other type of lattices. So we create a very smooth function that you change the variable, you get a different lattice. So this you can even mix and match lattices. That’s what’s happening here. It’s like a 50% star and 50% Voronoi. These words are more or less have the same meaning if you look at the literature.
5:12 We are just following e-literature because Kagome means slightly different thing depending upon who you ask. So you can now create new cells because there is a smooth parameterization, and that means you can populate a lattice like this. In each node is a different topology. Right? Okay. So we This is when I would say we created a monster. What is the monster? Monster is we gave infinitely many variable to choose a lattice from, and as this session’s talk have said, it’s pretty hard to choose from.
5:45 So it was like, “Okay, do simulations. So, we took a puck, we generated a mesh, we generated a lattice, and we created a lattice that is conformal, we put it at a hydro mesh inside it, and did a simulation. We tried ANSYS, Abaqus, Nastran. Our materials are non-linear, so neo-Hookean. Not a single one of the commercially available software can do the simulation except for one. It’s not commercially available, and for this I have to give Daniel A and group a huge shout-out.
6:15 There’s an algorithm called incremental potential contact, which was developed at NYU, and we started working with them. We’ve been working with them for past 2-3 years. This, for FEA nerds, this is implicit solver. The time step does not go down exponentially as the buckling and the contact happens. And this, once we figured out how to make it work, there were some parameter tunings, but tuning is done, it’s in the box.
6:41 These simulations can run for different lattice topologies without any human in the loop. Absolutely necessary if unless we would have to hire a simulation engineer for every time you want to generate a simulation. Still a problem. The simulations take hours to run. It’s a phenomenal method, extremely robust, always works, but hours to run. So, we still haven’t killed the monster. So, this is where, we started developing what I would call, machine learning for physics simulations.
7:10 But it’s pretty hard to generate the data. So, we kept ourselves to not blowing our AWS budget, and only have 200 data points. The the key here is creating a machine learning model that understands the physics. So, we added physical constraint of the problem that we know about, which is stress versus strain is semi-monotonic. You increase the strut diameter, stress is going to be higher, at least higher.
7:39 Cell size goes higher or volume fraction goes down, stress is going to be lower. So, you can keep adding these constraints that are more or less physical into it and use the off-the-shelf neural operators that are in the field for these type of approaches. And you can actually capture a pretty good response. So, the input is all of your lattice parameters, which cell types, topologies, strut diameter, volume fraction.
8:02 This is the strain of the time because these are quasi-static simulations. You do the multiplication and the output is the stress. That’s what we want to predict. We started with 200 data points. And of course, I’m going to show you the slides where it’s perfect match. Stealing the joke here. But You see this slide is just talking about where the cell type is the same. Now, we are looking at the trained data.
8:29 So, the model has not seen these type of this exact lattice parameters before. And then the blue and the red is matching. It’s not exactly matching because you see the Y axis has a pretty large numbers. So, the error is less than 3%. But it is capturing non-linear mechanical response because you will have a plateau, you will have a densification, contact, everything. This is the end result that we care about.
8:54 Now, you can take it to town. So, we can only run simulation for some lattices up to this point because the AWS budget had some limits. But the machine learning has no limits because it’s capturing and the lattice topologies are different. Of course, once it’s trained, it’s working. It’s extremely fast in running simulations even though our simulation method, whatever we chose to adopt, was pretty robust. Okay, this is pretty cool.
9:20 How does it affect the users? Within the design design engine, which is the product we created, we have a lattice library pretty similar to the concept that has been explained before in talks today. This is what I call kayak search of lattice. So, you can tell exactly where you want to go, the stress plateau, 25% strain, etc., etc. The type of material you want to choose. So, instead of the 9,000 data points that I’m sure we are going to continue to add more of it, you can sort of drill it down and get a whole mechanical response and choose the stuff that you need.
9:52 And you go to here. Say you want to populate a lattice, I create three different zones. You choose any of the stuff that you can pick from here and it will automatically populate. So, this is where all the boxes are lining up. Why we can choose from anything in 9,000? Because we know that they will always be compact and no hanging outs. The actual other part that we could actually even develop the machine learning model here, the other truth that came handy was we did automatic differentiation on the whole model.
10:19 So, it’s not a discrete variable optimization. And it’s possible because the parameters that go in the lattice topology are smooth. So, if they weren’t smooth, you wouldn’t have that. So, then you cannot use the standard optimization method and sort of the problem falls apart. But applications are always interesting. This is where I was starting the talk. So, we have helped CCM, which is one of our customers, create a helmet for ice hockey from the day we had the first conversation to within 7 months.
10:49 The The helmet is on the field and the customer cared about developing this helmet not because it’s cool and 3D printing. That’s why I would do it, but they cared about for the superior response, right? So, that’s the goal. And this is from day one a mass customized product. Each design is going to be different because head shape is different. We basically worked with the company called Toolkit 3D, which is now called ShapeShift ShapeShift 3D, now called Toolkit 3D.
11:18 They do the scanning and create the CAD and then we generate the lattice. We put the lattice on the print bed. This has been happening since 2019. Not a single human in the loop. And every time CCM places an order, it goes through the system. There are really complicated stuff. Some other problem that are not lattice related, but thought I would highlight one of them. This surface is a customer-facing surface, so it’s going to be unique to everybody’s head shape, and we put a texture on it.
11:46 So, of course, we had to solve a UV mapping problem. So, there are a lot of pieces underneath here that are beyond lattices that we had to make it robust enough so that every helmet that is going outside of the building, it performs the way it performs, and each helmet pad has several different lattice zones in it. Helmet is a regulated industry. So, the customer cannot just sell a customized product without a guarantee of it working the way it’s supposed to work.
12:11 So, we actually created algorithms where the shell side of the surface will have the same nodes at the exact same locations. We cared about we did the performance characterization of the new part that’s going out of the door, which is unique every time. We don’t want to run FEA, so it’s all heuristics and statistical modeling, now called ML. And we also look at volume fraction, so make sure that the part is printable, because if the part fails at the time of printing, it’s a yield, and somebody in the CM is going to shout at us.
12:37 So, it has to work both ways. So, by in this particular place, we are going all the way from data capture to CAD to simulation to print prep, sending it to CM in Taiwan without any human in the loop. Second product, with Vicis. There are several examples, I’m just highlighting a few. Vicis developed NFL helmets. They are the top-rated helmet right now in the safety standards. This is also a customized product.
13:04 NFL releases the poster of the safe safest helmet every year. If you look at the poster this year, there are two companies, Riddell and Vicis. They are both working with the same technology we have provided. Here are some quotes from them. From their perspective, the part that I think we have focused on, which is more and more simple. I don’t think it’s simple enough for mass adoption yes, yet but I hope we get there.
13:33 And because the demo will not be over if there is no chat GPT plugin. So, come on. You can email me if you have questions but clap after the demo, okay? So, this is our tool. It’s a cloud-based tool and we have this person called Wilson in our lab. He’s been doing lattice characterization for all his life and so we created a tool for him. In which you can say I need stress versus strain with whatever numbers you want, generate a lattice for you and you can basically, let’s just run it again.
14:26 I’m just making it up. There you go. Didn’t work. Of course. There you go. Okay, so it’s clearly not up today but you can get to the points where you let me show you the pre-recorded stuff that without the chat GPT, you we are able to get to the type of mechanical response that the stand the most foams have. So, we have characterized the foams. The data is already available, whether it’s Poron or EPP 60 g per liter and now you can just figure out which lattices you want to use when your base is a material that you already know what you’re using with because that’s the use case we’ve seen in the field quite a bit. In the incumbent there is a material and now you want to test against it. Cool. I’ll pause there. Are there any questions? Happy to answer.
Listen, search and cite
Listen to this talk as a podcast
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

Reinventing the Wheel – Representing & Manufacturing Geometry with a Program
Mark Burhop · Siemens Technology

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























