CDFAM NYC 2023 · New York · 14–15 June 2023
Generate-to-Validate: Optimal Designs and the Need to Deliver Under Uncertainty
Abstract
risk & uncertainty, robust design, stochastic optimization, CAE workflows, design‑for‑manufacturing, topology optimization, computational engineering, adaptive meshing, product reliability, additive + injection molding, structural optimization, engineering cost reduction, sustainable design, variability-aware modeling, rapid design iteration
Transcript
From YouTube’s automatic captions, lightly cleaned; expect some errors. Each timestamp opens the video at that moment.
Read the full transcript · 3,562 words
Foreign. I’m Andre, I’m the founder and CEO of Rafinex, and in the next 20 odd minutes I will try to talk about two things, mostly risk and speed, and the need to to validate the work that you’re doing. I will be fine, thank you very much. I will try to touch Sean as much points that we mentioned this morning, this morning, as possible, but costly, four things, four blocks I want to talk about, really quickly, who we are, because I think this is our first in presence, yes, appearance.
0:40 So just a quick word on who we are, where I come from. Then my big passion is in and around risk, risk management in engineering design, and how that enables or destroys sustainability and the profits that come with design optimization. Then we’ll actually dive a bit into how that actually works, I am not holding back, and how we do things. I think that’s very important, to also help with the language and the know-how in the community. And then at the very end, I’ll actually discuss this on the ground with a diamond, Dupont, a cup heart, and how that actually turns into money on the on the road.
1:20 Quick thing, so Refinex, I created a company in and around algorithmic innovation, so we really focus on a much more efficient, sustainable future. So building cars, fridges, anything, we cannot continue this resource hungry along the world. Our niche contribution to that kind of area is developing new CAE, computational analysis capabilities, and and build those into into software. We are a European outfit from there, from the get-go, I would say, so we have people from Berlin, we have people from London, myself, I was at Imperial for quite some time, and our knowledge is really HPC, FEM methods, numerics. Actually it was mentioned this morning, very few people know how to actually run an HPC with 10 000 cores, well I’ve got people at home to do that.
2:14 So, second part, why risk matters and how it is key, or not key, to sustainability at scale while making profits. So one thing that is often overlooked, or not exploited enough yet, is the fact that 70, 80 percent of products’ impact and costs throughout its lifetime are driven during the design phase, so unlike day one. So you can spend double digit millions on engineer, like machines and powder control and process and material development, all of that, but at the end of the day, if you happen to choose the wrong shape, you’re still in trouble, right? So this is one lever that we have to go off, go after really, really hard, if we want to build cars have to wait twice the range, that kind of thing, I’m not talking in incremental improvements.
3:03 And there’s this guy from the front, no, yeah, the front of Institute, who said that really lightweight and design and recycling all start on day zero, at the design phase. So this is really, this is where the value is, this is the carrot. Now, what does profitability at skill really means? It means building stuff using a lot less energy, a lot less materials, use less energy to the operator. That’s optimization tradecraft that has been around for some time. So why is it not mainstream? Our theory, or one key contribution to that, is it’s not fast enough, it doesn’t account for risk, because if your product has reliability issues, at the end of the day nobody will do that, right? So how do we do the first part while taking care of risk and reliability and being fast enough to do this on a daily basis? So if it’s risk that really kills the deal, so how are we doing that today, right?
4:01 So let’s say we’re designing some product, and there’s some requirements that we kind of openly know, we put them in new requirements envelope, which may or may not actually be complete, but we wholeheartedly know that there’s other off-design conditions that this product will see out there in the world, and that is a known risk, but we don’t, we know it’s there but we can’t quite quantify it. Well, there is ways to quantify then, the way we, but the problem is that when stuff starts moving, anything that moves, just making your design fatter in every direction is not necessarily a smart move, because it might actually aggravate certain conditions, especially when it moves and has to stop, in accelerating, detailing masses. So you only want to add mass in those directions that are relevant, while keeping irrelevant or safe directions on, like there’s no need to put extra mass there. So it’s this one key contributing fact is that wholesale application of factors of safety make or destroy optimization a bit from the get-go.
5:06 So that brings me to how, so how do we actually do that, how do we still optimize with kind of active knowledge of of risk management. And this is what we specialize in, so we are really specialists in stochastic optimization, particularly in context of topology optimization. And for the specialists in rumors, it’s a merger between UQ methods into classical topology optimization. Language is also touched upon a few times today already, I’m a big fan of the classics, if you haven’t invented anything new, don’t give it a new name, use what everybody else calls it, right? So no, no need to go hubris or confuse the market with hyperinflation of names.
5:45 Oh yeah, and then, so this is kind of a standard field that will operate for the next couple minutes. The only difference, or the key difference, for stochastic optimization, is that anything that has to do with inputs, can be loads, can be material properties, they don’t have to be assumed to be perfect. So, but your risk, your variabilities that you know or can estimate, they are baked into here, they’re baked into the inputs, and these variabilities can either be known because you managed to learn the lab, they can come from IoT data, I think there was a nice picture from the running shoe this morning, right, they can actually come back from sensored equipment, or they can be prescribed by legislation or so on and so forth.
6:21 So if you set up your optimization problem like this, this will happen. So if you make the optimization aware of variability, it will be able to sense in what directions this variability is important, and in what direction is not important. It’s key for that, that these variability conditions are integrated into the search gradient of the optimum, otherwise this trick doesn’t work. But what happens is that you don’t get idealized optimal points, but you get formally robust optimal points.
6:50 There’s a reason we’re sitting on shares and not sitting on Segways, for instance, this is kind of the logic. So there is, there are better designs out there, and there are idealized designs out there, for in terms of solution space and math surface, kind of thing. Standard optimization will always try to reach, locally speaking, the highest peak in the mountain range, versus robust optimization will try to find the highest plateau in that mountain range, so that if your design is perturbed from your assumed conditions in operations, it won’t suddenly suffer a massive crash in performance. So you get this kind of safety buffer around you, so that’s the whole, the whole concept.
7:28 Quick side note, because it’s actually important, we also use a fully adaptive Ted meshing doing optimization, which has several convergence and engineering quality properties, but more than anything else, we don’t need to do any filtering of oxillo meshing, or kind of smoothening, and the reconstruction later on is also much, much better. So this, this is key, why, because robust designs, they don’t look super Bugatti brake pedal crazy when you take them to the stand, but they look kind of reasonable, and they tend to have very good vibration properties. They can develop shell walls, because thus it’s due to the adaptive mashing, shell structures is something that you rarely see come out of a normal optimizer. And they, like you call it this morning, they just look right, like you can, as an engineer, you can look at them, and once they’re there you can look at them, and you can kind of follow why they do that job the way they do it.
8:25 Robust designs also have this neat feature that they are forgiving, meaning that you might not need the A plus plus grade of the material to build your design, because they have very smooth stress fields and stress behaviors within them, you might get away with using grade B material, so you can get, save a lot of costs there. So we have seen this in the past, so in that sense, I hope I have convinced you that bringing the awareness of real-life variability into CA, into optimization, leads to much more applicable, fit for reality, type designs.
8:58 Now, off to my second part of this, of the speech, which is validate. All of these claims I can say this, but none of it should, too often it’s not properly discussed or properly mentioned, but everything I just said here, from stiffness performance to strain energy performances, all of these are proper FEA, top 10 based, nastron style, let’s say, but it’s not NASCAR, and it’s our own code, but it’s equivalent validation.
9:21 So here we can see, effectively, this is the mountain range, we’re sitting on top of a flat mountain rather than a steep mountain, and also doing off design conditions, our strain energy doesn’t go through the roof. Ergo, this explains why we have better frequency and better eigen mode properties as well, because the two, these are linked, yeah, here we go. So robust designs have better frequency properties, which is key for automotive, but also aerospace and space applications.
9:52 Also, while everything I just said is effectively linear physics with uncertainty driven inputs and search gradients, and getting, let’s say, a statically robust designs, you have follow-on, rat tail, positive impacts on durability and frequency properties, because you have this advantage of, robust designs have this smooth stress field within them. So even you do a lot of good on day one if you do robust designs.
10:24 Also, it’s not all about additive, I mean, here are examples of straight outputs, so that was actually a robot part that picks up randomly oriented parts. So the question was, how do I optimize the robot gripper for randomly oriented parts? The load conditions are not properly defined, so we managed to get through that, and in the end I just hit print and put it on the robot, and it worked. This also exists in hollow, if you want it. But the key is, sometimes it’s said that generative design equal equal manufacturability, versus good old topology optimization can’t do that, that’s a lie, manufacturability constraints exist in existing topology optimization codes, so I want to clean up some misconfusion, confusion, there.
11:04 Yeah, and then, if you take away one thing from today, please judge designs on their numbers, and note, of that looks, too many times you get generative portfolio, 25 designs on the screen, and then you go like, oh, number 17 kind of looks cool, but who says number 17 might have the worst frequency property. So always run the validation and look at the numbers, we can quantify robust.
11:26 And so this is a part I can’t show, it was a baseline design, there was a normal optimizer, and there was robust optimized design, and then throughout the operational conditions we looked at how often we stay below a certain thrush threshold. You can see we did that for 95 of the cases for the robust design, as opposed to 77 for the baseline design. So we find you this advantage, and then you can do two things with that, you can say thank you very much for the extra safety, or you can trade in that extra safety, maybe 77 for your product was adequate, so you can then take mass out of the system until effectively you get back to your 77 percent robustness level.
12:08 Just a few pictures to to make it real, so we work a lot on automotive e powertrain holders, suspension arms, that kind of stuff. One very interesting field that we operate as well is, you can have variability occur from the real life, like real life load conditions fluctuate, but you have another area of uncertainty where it comes from, which is project management. Greenfield project with undefined requirements on day one, if you do standard optimization, you do a one optimization today, you change your assumptions tomorrow, and then your design may change completely, robust optimization doesn’t do that.
12:41 So as long as you can say EV tolls, who’s built EV tolls before, right, there’s no historical reference cases, so as long as you’re able to say, I can reasonably assume my loads to be between A and B, with robust optimization can get going, you can find good macro structures, good torsion, health stiffness designs, and then you can make project management decisions on that basis. So very real, very show me the money.
13:09 This is a part, it’s an engine mount from a Daimler car, it used to be metallic, pre-2017, 2018. So then, alright, king of dominance illness sadness is formerly duper switch this to a polymer magic Source Depot polymer, with some fibers in it, don’t ask me. Anyways, they went from metal to plastic with a new injection molding process, took them a few years to get that right, but it’s a production part. But then the question was, okay, then, yeah, so this is the switches zinto on the edge.
13:46 But then last year Seleness came to us and said, well, the injection molding process has so much more freedom, how about now we go after the design. So that was, that was very interesting, because for them it was a new design, or an application with a material type and production process which hasn’t been done in that form before. So we were tasked with showing them how fast we can be, like how fast is the turnaround time. They had a similar turnaround time, about a month per part, they needed, obviously, the same performance levels, and they needed it to still be injection moldable, that comes with member size constraint and that kind of stuff.
14:26 So I’ll quickly run through what our product, which I tried to avoid too much today talking about the way, is, maybe it does all of this automatically, in like four steps, and we are working hard on adding a fifth step. So it all starts here, so it’s a browser-based ISO clouds have cured kind of SAS product, we operate in defense and aerospace as well, and so you start with your model setup, so CAD colored CAD, or even orphan meshes, as starters.
14:55 So there’s also a lot of knowledge that goes into how to properly set up such models and boundary conditions, the devil is in the details, how to do clamped bolts, not clamp bolts, all these things, so a shout out to the MIT team on the education part. There is some art and some deep knowledge on how to properly model connections in FEM, and even optimization, because it does influence the design substantially. Off we go.
15:25 Second step, we do our robust optimization search algorithms together with our adaptive meshing, and out comes a design. Now we, everything I showed today, they are straight FEM meshes, so we, we run on adaptive tests, we don’t do filtering, we don’t do interpolation, we don’t do smoothing. So that means that once the optimization is done, we have a really high resolution part under the hood that we can just extract, and we change the meshes from linear to second order. Why? Because now we can do a proper stress based FVA analysis, and this is usually one of the hidden trip wires in the workflow, because people sit there fiddling with the cutoff thing, or the other alternative is to recreate inspiration based the design on the second screen, it wastes days.
16:12 This is where we preserve also a lot of the curvatures on the on the on the joint areas, where the frequency properties come from, we preserve all of that work that the optimizer has done. So this is fully automated, I don’t think anyone in out there does it like that. So here, in this case, we just ran three volume fraction examples. Off we go.
16:32 FVA validation, only abundant conditions information is maintained, and no re-encoding them into a sec, or jumping from CAD to topology, from topology to analysis codes. If you still need to do further national linear based analysis later on, we can just download all PDF files and shove them over. So in this case we fell down on, on the middle one, or the 15 volume fraction model, it met all the requirements, the member size, the draw directions, the draw angle, all that kind of stuff was, was there, and the numbers added up.
17:10 Now, Dupont and went away, and they did the Digimat non-linear material checks, we weren’t capable, we won’t be capable of doing that anytime soon. And they also did mold flow injection molding simulations, and we cleared those, so that now they’re actually looking at building test molds for this part. And they were kind enough to give PR exposure in Germany and in the US, actually last week, about this whole design, design, workflow.
17:40 So just, we’ve obnoxiously clear on conclusions, this workflow for Dupont, did it, used to be three hours, but we made the code twice as fast last month, so now it’s an hour and a half, this design, in an hour and a half, plus FEA validations, plus analysis and discussion and so on and so forth. They decided on this design in two and a half days compared to about 30 days before. We took out the 25 of the mass, the strain energy dropped by 30, the frequencies went up, and you, you still have this robust design around it. So, quite a big milestone success for us.
18:23 Because it was mentioned this morning, as any other post-processing kernels on them that either suppliers or customers give us, so here we’ve been working with Module Works out of Aachen for about four or five years. For those who don’t know, if you open up your CAM software, long enough, ninety percent of the market share, the tool path and G-Code generators, is Module Works. So we just went straight to the source, they gave us their PBF and one tool pass generators, so we can just run them, and then, if you were to be happy and have the machine, you can directly connect there. But this is prototype, we aim to get this up and running for Form Next in November.
18:57 And then, as well, on the non-linear, because you may have picked up on the words frequency, uncertainty, reliability, both from the APL and from NASA this morning, having the proper validation in-house of buckling, random vibration, transient non-linear responses. So I’m working with some defense customers that maybe in about a month’s time there should be news, news there, just because also somebody mentioned it to smoke, I remember.
19:25 We also run anisotropic codes, so, and yes, anisotropic carbon winding designs are different, so the straighter one in the middle is run with on the fly dynamic and isotropy, this is the one on the right assumes some isotropic materials. So we’re working on special flavor physics designs, this is, this is a lot of work, it’s been chaperoned by G Can and McLaren out of the UK.
19:54 So hopefully I’ve, I’ve convinced you that it is possible today to be profitable and sustainable at the same time, actually you should, and you want to go after it. Some customers, even again, know, cutting three grams, four grams of plastic out of a wish mop bucket, saves them a million in purchasing costs each year, that’s the kind of levels that, that we deal with. So it doesn’t always have to be the flashiest of designs, you can do a lot of good with small modifications as well, but this only works if you’re fast enough, and if people have the trust, the validated trust, that if you strap your robot gripper into the robot and send it to Hamburg for the Airbus Factory, that it won’t die on you.
20:39 So yeah, key takeaways for today, being aware of uncertainty, being aware of risk, bringing those things into the optimization, is absolutely crucial. The math is not necessarily the simplest, but you will get better designs, and you will get them a lot faster. And that’s me, how was I on time, let’s put on, good, thanks.
More from André Wilmes
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

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

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
























