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

Design Automation Using Machine Learning & Computational Design

Listen as a podcast

Abstract

This recording is part of the CDFAM Computational Design (+DfAM) Symposium and features Rhushik Matroja, CEO of Cognitive Design Systems.

Presentation Title: Design Automation Using Machine Learning & Computational Design

In this presentation, Rhushik Matroja, co-founder of Cognitive Design Systems, explores the intersection of machine learning, computational design, and manufacturing. Based in Toulouse, France, Cognitive Design Systems focuses on bridging the gap between concept and manufacturable products. Key highlights include:

The Problem with Concept Cars: Matroja discusses the disconnect between concept designs and what is actually manufacturable, emphasizing the need to consider performance, cost, manufacturability, and carbon footprint.

Automating Key Performance Indicators (KPIs): He introduces four KPIs that, if automated, can accelerate the process of bringing products to market: performance, manufacturability, cost, and carbon footprint.

Manufacturing Analysis: Matroja explains how Cognitive Design Systems analyzes 3D models using mesh, voxel, and implicit modeling technologies, considering various manufacturing constraints for molding, milling, and additive manufacturing.

Geometric Engine: He describes the company’s geometric engine, which uses mesh, voxel, and implicit modeling to automate design, focusing on manufacturability, cost, and carbon footprint.

Transcript

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

Read the full transcript · 2,803 words

Foreign. Thank you very much everybody, let’s talk about design automation using machine learning, computational design, to make products manufacturable. Chelsea, maybe I have few things in our tools that you might want it, so you see, I think it was programmed perfectly well. A little bit about about the background, about who I am, where we come from, so we are cognitive Design Systems, we are based in Toulouse, south of France, we were born two years ago out of the the need of whatever the Chelsea mentioned, how to bring the concept into a manufacturable product, but not only manufacturable product, a marketable product, because that’s something that is very key and it’s not easily achieved.

1:04 Currently we work a lot with with our Japanese customers, because our, my past is from Japan, I used to do a lot of printing and failing the printing, mainly for energy manufacturing, qualifying parts for Aerospace industry, medical industry, automotive industry, applying various different Quality Systems. So that’s where we learned that okay there is a big bridge that we need to connect between what’s being designed and what’s being manufactured. So the problem that here we are trying to solve is that, have you ever seen the concept cars that you see in the trade shows on the street? Well no, because you change everything with the performance analysis, manufacturable analysis, cost analysis, carbon footprint analysis, and this takes thousands of hours of design process.

1:59 So our solution is a little bit different, we have analyzed that there are four kpis, if you are able to automate them and accelerate them, we are able to maybe have some products which are ready to go to the market in the shortest time possible. Those kpis are performance kpis, manufacturability apis, cost, and carbon footprint. We have a lot of things, a lot of tools and solutions available already in the market which are awesome, who does the Fe analysis, mechanical, thermal, vibrational analysis, optimization, on top of that, it’s it’s good, it’s amazing, but we need to focus, as mentioned by my previous, if you don’t match the cost it won’t, the product will never be on the market. And more recent access is the carbon footprint, so I added just some, a sentence that team said just before, just because it’s cool it won’t be on the plane, right, so we need to make it happen, and in order to do that we need to analyze and optimize all these four kpis.

3:14 Of course the performance kpis, many many of my colleagues are doing it perfectly well, so we have kind of left it open to to them, and we are taking care of the three different kpis, manufacturability, costs and carbon footprint, analyze them, assess them, and then optimize them, and we are connecting our solutions to other softwares, plugins, to bring those added values to the cad environment, so the designer can have insights from the manufacturing plant right in his CAD environment, so he can design better from the scratch.

3:56 For that, with other two main categories of Technologies, the first one being the manufacturing analysis, so we analyze Parts, the 3D models, using mesh, Rock cell, implicit modeling Technologies, there are a lot of manufacturing analysis going on there, thin wall structures, thermal concentration, surface roughness, layer delamination, but also other manufacturing constraints for molding and Milling, also those are also very important. And the the second, second part of the technology is the our geometric engine, which is a also the mesh, voxel and implicit modeling based geometric engine, nothing new, a lot of people are doing already there, but we don’t want to do it just to create lattice structures, no no offense there, it’s just we wanted to to to bring something out of it which can be useful to design or automate the design.

4:53 So that’s how we use it, first of all, we we analyze the path for manufacturing preparation, manufacturably analysis, quality analysis, serial production, cost analysis, and carbon filter analysis. There’s a lot of things to tell us, and this can be kind of overwhelming if you are doing a day-to-day job, that’s what I was doing, I was actually assessing those parts that were, that I was receiving by the customers, in hundreds, and by one by one by one part we had to assess that, okay, how can I print this part, if I have to print in this orientation, what is the the manufacturability looks like, will direct me deformation, there will, there will, there be any any features which will be distorted, heat concentration etc, where the surface roughness will be rougher, smoother, where do I need to add some over thickness to remove those those zones. Those kind of analysis were done manually, takes hours and hours and days of of work.

6:02 So what we we wanted to do is that with my colleague Ari, who is also the CTO of my company, we wanted to go to Hawaii by automating this thing, and in order to do that we created a software called cognitive additive, which allows to do that automatically, you can assess those those insights in few seconds. So you will be able to focus on on how to to bring those, how can, how to assess those kpis and design better. Why we wanted to to do it in in such a short time, it’s because these are the the KPS that we’ll be optimizing, and to optimize them it’s not only topology optimization, that’s one way of optimizing path, but as Chelsea mentioned about amfam, we need to optimize locally areas where which has some issues.

6:54 So we developed a result, our solvers on simulated alinee, link conjugated gradient, proximal gradient etc, and we optimize for manufacturability, we optimize for cost, modify the geometry documents for cost, and we are working on optimizing it for carbon footprint, so that’s our goal. Of course we use a lot of field driven design, because that that enabled us to to modify those geometries, then we have of course mesh and voxel modeling, which allows us to, there are some certain problems that could be assessed by by only modifying some mesh, we do that, or by voxel we do that also, and we use also implicit modeling when we, when it’s required to to put some complex structures there, or doing some Boolean operations, smoothing etc.

7:49 Sorry, I forgot to click, but yeah, here you can see the part that we can, we can morph using our one of our tool which is made for molding, so you can see that the part which is, which is imported, it’s a it’s a part which is, it’s topology optimized, but we wanted to make sure that okay this part has to be moldable, so we we created a tool which actually assess the the molding analysis, it’s not just additive that we are talking about here, also the molding, and we modify the geometry, we we completely change the geometry to make it, make it more moldable, and of course while doing that we also create the molds itself, so we are able to assess the feasibility of the mold, because mold, if it’s made through additive, it’s made through Milling, it won’t look same, cooling channels inside won’t be same, so we need to to to get those those inputs outside, and for that purpose we have a tool that is also present as an SDK that I will talk about later.

8:51 This is a representation of that SDK, energy manufacturing, specifically how we do is we analyze geometry, we assess for each component, we we bring out also a dashboard, which is easy to communicate internally inside the the company with your various colleagues, to to Showcase them that okay in a colored map here is the zone that you need to modify, if you are not the designer you need to ask the designer to to change that, and then you can also automate that, that’s the next step, by doing this local modification, so we are able to to do apply those modifications autonomously, converting cylinders into teardrops, changing some of the the overhang angles to to reduce the support structures required, adding some over thicknesses where the the surface roughness is very very difficult, etc, etc, and then we evolved the design, we we assessed the the, we assess those each design with four, four criterias that I mentioned, performance, cost, manufacturability and carbon footprint, and we we put it together in a parietal graph, and then we can see that okay how my part is being evolved and how can I go to the the minimal and find the the the global Minima.

10:14 The tool that we have for that is cognitive additive, it’s a tool which is which is already out there, you can you can you can get a demo if you want, don’t hesitate to ask me. This Tool allows us to, you to assess the manufacturability analysis, cost analysis, you can automate the orientation and nesting, you can do path screening also, so we we took that that analysis and we brought it to the large scale, so you can put your whole database in it and we will assess it very very easily, everything is on premise, for obvious reasons our customers are in Japan, are in France, not so open for the Cloud yet, so we had to create a tool, and we wanted to make it faster, because if we don’t create a tool which is, I mean, if it’s a very slow, then nobody will use it, so we accelerated it using various machine learning algorithms.

11:09 So that’s where the machine learning comes into play, it’s a it’s not a magic thing, it’s, we just, we have some algorithm items that we need to accelerate and we train models on, based on that. One example that we did with with a French customer, they make doors for the aircraft, they gave us with a with a sub assembly of 840 metal parts, this is the whole, it’s a huge file, and they wanted to see that okay what are the parts suitable for additive, and if those parts are suitable what’s the best machine to print with them. So finding those, if you do it manually, it takes months of job, so with our solution we were able to to to map it on on on a partograph with a feasibility cost, and then filter it down to the the 10 components which are actually making sense, because others were not making sense.

So yeah, those those calculations were done on 61 000 different solutions, so each component is studied on 10 different powder bed Fusion machines, each powerbet Fusion machines has a various different print parameters and materials associated with it, so it becomes a very very large number of calculation that that needs to be done, it was done under seven hours of calculation, and we took way more time to analyze them, those data, because it was a lot of data.

12:35 On the molding, what we do it’s quite similar to additive, we analyze the geometry, we do the the automated dfm, and then we generate the molds itself. How we do it is by analyzing those geometries, we know where to modify those geometries, how to how to to to play with the those draft angles, those the wall thicknesses etc, and we will be also assessing that okay if I have to to mold or demold this part, will it be possible, how can I modify the mode itself. So this is a the loop which goes on and we have this kind of generation.

13:14 Which is done, an example that we did with Nissan in Japan, it’s, they came with us when Tesla brought out this Giga mold, that okay we have, we also want to do that, okay let’s go for it, they had a rough concept of what they want to do. So what we did is we did a structural optimization based on that, then we applied the cognitive molding, where we were able to completely morph the geometry, create the modes also, and we provided them, they said okay here you just gained us nine months of work, because the engineers who are doing sheet metal, Engineers who are doing molding, they don’t talk with each other, they are not not the same skills there, so we needed to to to connect that.

13:59 So you might have got the revision also on that, that we are not only focused on additive, we are going towards molding and machining in in coming months. A lot of comments I received on this this post that Juan posted, this is just an example of showing you what is feasible, you need to combine those those different methods of manufacturing in order to achieve the best part, it’s not one process which will give you the end product as it is.

14:27 So that’s our main work there, on the machine learning side, we have three strategies that we are using currently. First is that to accelerate some analysis using a n, it’s pretty pretty basic, I would say everybody is doing that, and we are we are also training our models to to to make it as accurate as possible with a lot of data sets that we have, with the NYU ABC data set and another data set, but we are also doing a lot of work on the feature detection, on the mesh, on implicit, that’s key, because once we know where the where the features are on those those geometries, we are able to modify them, but also reconstruct them as a b rep, and we are all able to bring it out to the to the cad environment to continue the process, because once it’s mesh, if there is no go, then it the part is useless.

Also third one is the the the generative adversarial networks, where we are just playing around, it’s it’s very beginning of, it’s in infancy yet, we are able to generate designs where it’s designed to modify them locally, and we assessed those manufacturability costs for each one of them to to make it happen. Secondly, the the computational design, so that’s the tool that you saw a little bit earlier also, that’s a GUI of our SDK, so we we provide the SDK to to whoever wants, and those SDK allows us to to actually play around, create your own workflows, once the workflow is set up, set it up, we can create a UI out of it, and you can you can share with your colleagues, share with your customers, that’s possible.

16:11 That’s what we are doing here, in the example we see that we we did a feasibility analysis, we saw that there are some small structures, thin thin layers, we we did the modification locally, we applied some variable thicknesses there, we did the simulation again, and it’s it’s, it has been cleared, so this is simple workflow that we can create, much more complex workflows, and that is not easy to present in a slideshow. We search some applications, so our very specific use cases, such as in defense, in in in medical, where we need Mass customization, and that’s where we we come in, we create our workflows and we accelerate them.

16:56 Here today, we are looking for Partnerships, so we we want to collaborate with with with the industry, a lot of people with Cloud environments where we can bring, we believe we can bring some value there, also CE environment, we already have some examples that we are already working with, for example CA, we are working with synera, where they have their own the platform, where we have, they have also cognitive attitude inside, so they are able to do the optimization Loop for topology optimization based on cost, that’s something which is which can be explored by you guys, if just, there is also somebody from Cinema here, give him a shout out for that.

17:35 And machine makers, we we have, we have partnered with with statuses for the first time to bring this feasibility and cost analysis to the to the larger audience. For that purpose, thank you very much for your time and attention, if you have any questions don’t hesitate to to ping me on on my email or on the on the website. If you want manufacturable and marketable products, we are here to help you for that, thank you. Foreign.

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

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.