0:00 Thanks for organizing, it’s pretty splendid place. My name is Andre, I’m from Rafinex, I’m one of the co-founders, the other co-founders in the audience at the back, Johannes. I want to put the focus today on risk and money, computational design, to get it into mainstream adoption. So this won’t be a product presentation, my product is at the back, so if you want a talk product we can do that back there. This is going to be technology and ideas about robust safety, how to get computational designs in in the air.

0:30 Four sections about what I want to talk about: first, who we are. I think we know a fair amount of people in the room, but not everyone quite yet, so just quick words for where we come from. Then I want to touch on, yeah, the role of risk and money about computational design in industry, on the field out there, up in space, and so forth, a bit about the theory, or at least one take, how we go about about it. And then I really want to move as far as possible forward, to as much as I can show about past, current, and future design activities that we involved in, to help promote computational design, bring more vehicles into the industry, to to really be a trailblazer in that sense.

1:15 So I put this together this morning. So Rafinex is actually a Berliner, so we are formerly spin out of, down the road from the VAV Institute of Technology, sorry, V Institute of Applied Analysis and Stochastics. So yeah, unfortunately, I thought initially that it was a good historical reference, but then today’s macro politics maybe brought us a little bit back to that situation as well. So why does risk matter, really? So we think that risk is integral when business decisions are being made, what designs are being produced. So if you don’t manage risk, then optimize, in, within the context of optimization, then you’re never really going to have the trust of the business owners or the design owners to commit to optimization. So we really maintain that you have to have that concept of risk and robustness when you go about really light waiting parts, saving money, saving raw material, saving fuel.

2:13 So where does this all come from? People always say good design, I don’t need to preach to, they converted here, that good design is worth it. But just how much worth is it? So the kind of life cycle footprint, environmental impact, or cost, order of magnitude 70 to 80% of the design are fixed during the design phase. So it’s well worth spending more time in design, because you reap the benefits down the road. But these are just about putting numbers out there to convince the wider audience, and not just us. And so are just some references. And that design is, I think, critical, good design is critical in saving materials at a large, large scale, because otherwise we can’t keep up with consumption. So if we really want to align the interest of sustainability with profitability, which is totally doable thir days, we have to do optimization, do more with less, be more efficient, and so on and so forth.

3:14 So the first part, the first three points, are effectively optimization tradecraft. A whole plethora of words have come across the industry recently, but this is new approaches, but at the end of the day is optimization tradecraft, this has been around. So that bothers me, the question is why isn’t it in daily usage, because optimization is not in daily mainstream usage. Our take on it, that one part of the problem is that the optimization is not fast enough, and it doesn’t account for this risk. Like, what’s my risk if I optimize, if I really take this much weight out of a design, what’s my business risk, am I going to risk recalls later down the road, yes or no? No one’s ever been fired for buying IBM kind of thinking. So we think that you have to actively address risk to create trust, to then create adoption.

4:01 So what’s the current solution? So risk through factors of safety can be a bit of a gund in optimization, because just scaling your loads in all directions and making your designs thicker in all directions, based on some scaling factors, can be counterintuitive and can be counterproductive. In the context of op of optimization, you might be spending materials in directions where, from where there is no risk, or you might make your own situation worse doing dynamic effects as well. So more is not always necessarily better, more in the right locations, yes, more in just about any location, not necessarily.

4:49 So how do we go about tackling, or making sure that we get optim optimal designs which are also robust? So this is where our take on it, our approach is is stochastic topology optimization. So for those not so familiar, it’s this idea of uncertainty quantification and merging it with PD and and and and search, gradient based search.

5:09 So here I’ll take a short pause in terms of terminology. I’ve spoken to a few people in the room, really, my my ask for the whole community, and this is a responsibility shared by us as well, is keep it simple. Like, if there is existing terminology for technologies or approaches, use them, do not make up new ones, it just confuses everyone. Like, tech taxonomy is important. So please just keep keep your marketing departments at bas sometimes, because otherwise we’re just destroying value for everyone in the market out there. So sorry, I had to take up that fight.

5:52 So we are within classical topology optimization, gradient search based, and the whole lot. The only thing where we do different is that our input, initial conditions, input data, particularly the loads and the load angles, instead of being single valued, perfectly known, we can associate a probabilistic distribution from on that, you can either estimate it to the best of your abilities, or maybe you you’ve measured it, or you can estimate it as well. But you are kind of embracing the fact that your input data isn’t fully known.

6:20 And so something really interesting happens, this is a bit of a simplified Lego example, everybody gets the point across quite well. Under idealized conditions you get idealized designs, but the world isn’t ideal. So there’s a reason the Eiffel Tower has four legs, and there’s a reason chairs have four legs, and there’s a reason why a segue without active control can stand up right. So that’s the idea of of robustness, and this is what we’re trying to bake into the geometries of the of the parts that we create.

6:49 In terms of mass optimization solution space language, effectively what we’re chasing is not the highest possible solution point on the surface, but we’re looking, actively pinging and actively searching for flat plateaus. So getting designs which have a high performance during nominal values, but should conditions, should assumptions be wrong, you’re moving around on a performance plateau which is fairly flat.

7:18 Another technology brick that we use, and that we do talk about, is we are running on adaptive meshing during optimization, automated t for meshing, doing optimization, which has a nice effect on the optimization, because we catch much smoother surfaces, we don’t have to geometrically filter. The de optimizer doesn’t suffer from numerical noise on surfaces and stuff like this, and it allows us also to create, for instance, shear walls doing optimization as well. So it’s it’s quite a nice Lego brick under the hood, which enables a lot of stuff for for us during during optimization.

7:54 Now, also, again trying to be a bit pragmatic, what we’ve seen over the last six years since we started, robust designs we call them in the house, ambitious but not totally crazy. So a priori you don’t necessarily know what the design will be, but atog, if you look at them, they make sense. Like, from a mechanical point of view, we often see common shapes, we see stuff that looks like U channels, we see stuff that looks like eye beams. So as a mechanical engineer, at hoc you look at the loads, you look at the design, and you go like, yeah, I understand how this thing does its job. And that’s for us is integral to creating trust as well into those geometries, and then thereby have adoption as well. So it’s not about being funky and looking like a coral reef just because you want to look like a coral leaf.

8:43 The other thing is these parts have, so these parts have this inherit robustness property, so they are lenient, they are forgiving against fluctuations, they’re also give you additional room to play against variabilities coming from manufacturing. So if your overall macro geometry is forgiving and robust, it means you can have more variability coming from the material side, for instance. So you can use powders with more recycling in there. I used to work in the ceramic industry, so that was always a big question, is like how much secondary ceramic tungsten powder can we add to to to the to the powder to make the new parts, what’s the risk, what’s the reliability of the part.

9:28 Now, again, back to numbers, our optimization runs in linear elements, that’s not not a secret, we run on the adaptive meshing, the optimizer finishes, and then we extract the shapes. I’ve heard some things this morning about disconnected regions and redesigning to inspired designs, we try to bypass that as much as we can. We we put a lot of effort into the meshing, so that the shapes always, let’s say, continuous, although it’s not a proper really word, but they they are always well defined, they water watertight, and so on. And then we extract that shape, we convert the mesh to a t 10 mesh, and then we do a proper FAA validation.

So what we can then do is, these are also, again, simplistic examples, but they get the point across, well, is designs with increasing variability during optimization have increasingly flat performance plateau in terms of stiffness. But we can also show it in the strain energy part, which is related to the stresses in your body. So these parts have this intrinsic robustness, but please always do this in FA validation, don’t judge parts because they look funky or whatever, do the FAA validation, look at the numbers.

10:47 This is something as well that I want to mention, although we can mathematically not guarantee it, but there is a good reason why there’s a correlation between robustness and higher first igen frequency. Although we cannot a priori guarantee it, typically we do see that robust designs have good first EG mode properties, which is good for NBH and and and other applications. Now, I won’t dwell into this topic, because this is a whole range of itself, but as far as we can tell with our customers is that manufacturing constraints are absolutely key. We are talking about designs which are formally robust, which are destined for operations. Ergo, unit costs matter, cost matters. So manufacturing constraints is a huge topic at the company, I won’t focus on them today, but they are critically key for for mainstream adoption.

11:35 And then this is my second big ask today, is, yes please, generative design, whatever computational design, judge designs on their validation numbers, don’t pick geometry number 17 just because it looks cool. Please trust the numbers, run the numbers, make fast and informed decisions, which is a slogan that we have in house as well, but the gentleman from Moon Rabbit this morning also mentioned it. It’s about generating shapes which are trustworthy, fast, doing the FA validation, and then formed the quantitive basis for people to make fast, informed decisions.

12:16 We can quantify robustness as well. So the same way we can stochastically optimize, we can also stochastically analyze, which means we get outputs of distributions of stress, distributions of def deformation. And what we can then do, for instance, in this example, we can show for X percentage of to be expected operating conditions we stay below a certain threshold. So if we do robust optimization, we can tend to improve on that, and then this robustness budget you can spend it in several different fashions. One, you can say, thank you very much, for the same mass I got more safety, that’s great. Maybe you say, well, my baseline design safety level was okay, please give me material back out, because I 77% reli, like my my 77% cut off was good enough. So in that way you can trade the the robustness budget back for for weight, or you can say, like we saw earlier, you can say, well, maybe now I downgrade my material to more recyclable material or a cheaper material or something like that. So you get to spend that budget whichever way you want.

13:27 Just moving a bit more into geometries, really, so we do two parts of two types of optimizations mostly. One is what we call high resolution near final, final designs, really high resolution parts that we try to get as far to the manufacturing state as possible, I’m not saying that we get you immediately finished designs either. The other one is also interesting, is we can, we can use use the stochastic inputs to not account for variability in the real life, but uncertainty coming from project management. Day one your requirements list might be patchy, semi undefined, or it has holes in it. So you can use this, the accessibility of having stochastic inputs, to at least get going, so it gets you over the hurdle of optimizing a first iteration robust design, which is least likely or less likely to dramatically change should you update your requirements later on. So we call this like fuzzy requirements optimization.

14:34 For those people coming to Euroset in Paris, this is what we are doing for a really funky Luxembourg company that does in the air carbon fiber, or 3D printing really large. There will be a car sized inspirational chassis on the stand there as well, so do do come by.

14:47 So actual designs, so in the past, so there also designs over there, more details over there as well. So this is a design undertaking we did with Dupo at the time, now SES, out of Geneva. They switched an engine mount bracket from metal to plastic injection molding in 2015, 16, 17, and then in 20, late 2022 they approached us and said, well, what could you do about geometry. So we got the freedom to go after the geometry really hard, and try to find how much weight we can remove, and also in terms of speeding up their processes. So, what I mentioned earlier, the ability to go from optimization to validation straight away, without having the human recad or redesign in in the middle.

15:03 So this is the gentleman that we worked with at the time, very, very kind person, so can only recommend him. So he was really interested in how can also selz potentially quote for parts quicker, how can tier one and tier tws quickly get to a reasonable design without actually having the full requirements as well, and then do a quote on base of that and stay competitive that way. So this was also a bit of the thinking in there.

16:11 So just in terms of workflow, really quick, these are the only pictures of the UI will show. So we start with a design space, typical workflow, we apply boundary conditions. What we then do is we do our adaptive meshing with the stochastics, we then jump into FA validation on the basis of t 10 elements. We can run different variants, so we can kick off different mass volume fractions to get a range of of of designs, and then on the basis of the FAA results we can then pick whichever winner we prefer. And then, so this was effectively then the the 15% mass target version that was selected, that’s also at the back. And then after that they took this design as well, and they sent it through digimat and mold flow, and so on and so forth, and it it survived all of the the manufacturing tests as well. So that was quite good.

17:05 So at the bottom line, effectively we managed to shave off about 25% of the weight, get the strain energy down as well by about 30%. % And in terms of compute speed, this is still 1.5 million elements roughly, it runs in about 90 minutes. But the setup to decision time with several variants is now two and a half, three, maybe four days, when before it was order of magnitude 3 weeks. So that’s the, goes back to the fast informed decision making cycle.

17:42 So now, as of compared to last time we met in New York, now we’ve moved on to systems, assemblies, so springs, couplings, being able to optimize really large structures as well. So this is where we really exploit also our, or people say cloud first, but I, maybe server first approach is better, because server first still gives you the option of WR running the servers on prem or in the cloud, like, or public cloud, but let’s say server based approach. We can create all these asynchronous workflow, so the meshing is being done asynchronously to the optimization, everything is all like task based, and you don’t bottl NE the the user in the process. You can make sure that your software is adapted to the CPR architecture that you want, and vice versa, which is also very important.

18:31 And from a business model point of view as well, is we are also a house based business, predominantly efficiency cuts into my gross margin. So with this mod business model, and me not having a variable pricing component, it makes, has makes the point that my developers have every interest of creating efficient code, because I’m spending the money for the server computer, there you go. So, but that’s, that that’s areas where we’re working on now. So we here as well, oh yeah, our our wingland doesn’t look like the one from the Voronoi this morning, so our winglet optimizations go really, really large. So we are talking meter sized objects with a few millimeters, or in terms of feature size, that is possible because of the efficient coating scalability paired with the adequate hardware under under the hood.

19:21 Recently as well, trying, and this goes a bit in the direction of, scenario now coupling with the ecosystem, which was also mentioned, I think, by Mr Lker. I think play nice with others, integrate, create, provide APIs, create value that way. So now we are able to exchange load data and and input data from from multibody systems simulations, which then actually leads me to a scenario which is not public, but maybe order of magnitude days away, maybe we’ll see. So yeah, we we we are now providing APIs into scenar as well, where you can do all these wonderful reconstruction work as well, if you if you wanted to.

20:08 Finally, now, upcoming projects, and this is also maybe my third ask of today. I have several projects in the pipeline, I would really ask people who want to design, build, and slf fly components to approach me. I will try the next year or two to work together with partners, software partners, as well as manufacturers, as well as OEM and customers, in designing actual demonstrator systems, subsystems, vehicles. So we’re talking about drones in the sky, we’re talking about jet parts, we are talking about UNM ground vehicles, we’re talking about these things. So I really want to work together with everyone to be able to be able to go, for instance, here, and then have very low earth orbit satellites stood there in the middle of the design. So showing not just ourselves but also the wider world what good computational design looks like, and kind of enticing and growing the community that way.

21:11 So yeah, we work on on on additive manufacturing for large scale parts, like meter sized ceramic additive manufacturing for ultra stiff optical benches. We are working also on multimaterial and anisotropic optimiz ation, so a shout out to the gentlemen from Endless Industries for printing what we come up with, so thank you very much for that. So those are also R&D topics currently ongoing. And then finally, this is already mentioned, so highly an isotropic topology optimization for really, really funky manufacturing processes, we’re working on on those as well, ultralight structures, think satellites, aircraft parts, aircraft interior, those things. So this brings me back to my key message, which means if we want really computational design to go more mainstream, I hope I made a case for trust, I made the case for robustness. So that’s kind of my key takeaway, or I hope that’s the key takeaway of today, because if we don’t make it profitable, if we don’t show me the money, then optimization is always going to be a fringe, fringe operation.

So there’s Johanes in the back as well, there’s myself, come talk to us about anything from straightup commercial all the way to how we can help with education and research, on in the academic institutions we are open on on quite a few different topics. And yeah, for those chaps who are attending any of those events, you’ll find us exhibiting there as well. Thank you. Great, thanks.