0:00 Good, oh yeah, cool, good morning everyone. My name is Gabrielis, I come from Uni of Edinburgh, and I’ll talk to you this morning about sort of where we could use, and where we couldn’t or shouldn’t use AI in computational design. Sort of give a bit of an overview, a bit of my own perspective, where I think we could use it, where we couldn’t use it. So I’ll start with a slide about myself. I come from mang background, I did my PhD in design optimization, AI, and mechanical metal materials, so sort of very much along the lines of what a lot of PE you people do here.

0:38 And now I’m working at the University, looking at how we could use sort of computational methods and apply AI in energy and energy material sectors, sort of modeling energy structures and materials. So, and today, I’ll I’ll just start about talking about mechanical metal materials, because it’s CAM, and we all love lates. And then I move into how similar concept can be scaled up, and how can we design wind turbine blades. And then move on to surate models, and sort of last slides about find, sort of discussing what what Danan mentioned, can AI solve our problems, or can it not, or more like where can it work and where it can’t work. Cool.

1:28 So, because it’s 9:00 a.m. in the morning, and you’re all geometri geeks, I thought, why not over complicate things and talk about four dimensional poopes? Good morning. So they are these very interesting structures, at least to me, that have geometrical similarities with with fractal structures. So if you look top right, you probably see, you you see that nice structure there, that’s a fractal structure. Basically the beauty of it is, if you zoom into it, it has geometrical repetition with every zoom in level, and for poopes, four dimensional poopes actually have some very similar properties to that. So, just a bit demystified, you probably heard about a tesseract, that’s what you see there on the right, and sort of how it is built up.

2:22 It’s, if you look in one dimension, you would just have a line, in two dimension, you would have a square, three dimension, you would have a cube, and in four dimension, you would have a t tesseract. The only problem is, well, you can keep on going like this, because these are sort of mathematically defined geometrical structures, and you can have fifth dimension, six dimensional, and so on so forth. So that’s all the mathematical theer that I’m going to give you today. But basically the issue here is, is a four-dimensional structure, we live in three dimensions, the closest to the fourth dimension that we have is time. We can say that time is the fourth dimension, and we just sort of, you know, represent that four dimension as time, and you get something like this. You get your four poopes as projection in threedimensional space with time.

3:17 So if I now tell you that, you know, I could use these to make met material unit cells, or like lce unit cells, you can probably see that, you know, I can take a snapshot and create luses for each rotation of the structure. And, going that we’re at CDAM, sort of, what would you do next? You would build, build that up, you would build simulations, right? You would build elop plastic fracture mechanics models. In this case, you would run thousands of simulations to sort of try to see what would be mechanical properties, or where where we could, where we could go with that. And in this case, I sort of settled down on a optimization objective to maximiz stiffness while minimizing weight, so basically lightweighting the structure. And of course, as most of you do it here, the next stage is manufacturing, so 3D printed.

4:17 And basically we had a develop sets of these meth material structures for comprension, compression, tension, and shear, and if you want to know more about it, their publication at top, perfect plug for my research. And then, of course, you know, you do mechanical characterization, you see how it fits with other structures. And the only thing that I would like to get your attention on, you know, like a lot of people do gyroid, and then you ask them why do you use gyroid, they don’t know. So if you’re using gyroid for mechanical properties such as stiffness or strength, this is where gyroid sits, this is where all the other structure it, I mean, again, I’m I’m not saying go and use these structures, but think about what you’re using.

5:06 So, but basically, this problem that I described to you, you generate a structure, run simulations, run optimization, and so on, can be, that optimization problem can be defined as a looking for a petto front, especially if it’s a multiobjective optimization problem, right? So I said stiffness and lightweighting, so basically you explore your design space with different variations of the structure, then you draw your petto front, and then you look for the point that is closest to your ideal point. This is, this is just basic optimization, right, and you know, you do, I don’t know, genetic algorithm for that, or particles form, or whatever you want to use, but but you get the idea, right.

5:53 So idea, the idea of this, this is your, like, forward design problem, right, you start with some design parameters, and then you sort of get through your simulations, optimization, to get a rage your performance measures, right, and it’s very good, and and it’s, but it’s sort of very computationally intense. So I sort of asked myself this question, could we do it otherwise, could we do it the other way around, could we do the inverse problem, sort of tell your framework that these are the properties that I want, and ask it to sort of generate structures that best Sue that type of application. And this is where actually my sort of dive into AI, ML, and at the moment NLP, sort of started. I sort of went in pro literature, and sort of a little bit like Tweed out what works, what doesn’t work, what people use, cuz inverse design problems are not new, people are using it, but they’re different ways to approach it.

6:57 So yeah, looked at anything that falls in the umbrella of artificial intelligence for design, specifically of mechanical metal materials and laes. If, again, another plug, if you want to know more about it, there’s a paper, the only thing that this came out in February, and we were talking with Danan, you know, like maybe we could sort of, maybe I could present on this. Since then, there’s a new sort of trend of using large language models for optimiz iation, so this is basically what your Chad GBT runs on, so imagine when you’re putting your prompt in, imagine that as an input for your optimization problem. So again, if you’re interested in this, I’ll encourage read my paper, and then look, look at that stuff, because that’s groundbreaking at the moment, I think. Okay.

7:46 And now about the big structures. So this was all work on mechanical meth material, small scale. And then, at the same time, we had company approaches, Tough Composites, really brilliant guys there, they work on sort of composite materials, renewable energy applications. And that moment, they were looking at how could you redesign a wind turbine blade, and their problem was, was this, they were saying, could you make it lighter? And why do you want to make wind turbine blade lighter? It’s it’s pretty straightforward, they’re getting bigger, bigger with every air, with every iteration, and because they’re spinning around the hub, one of the main loading conditions there is, well, one of the main failure modes are fatig loading. So the higher the weight of the blade, the higher the fatig loading, and, you know, the shorter the lifespan. So if you could actually reduce the weight, you you could, you could make them last longer, theoretically.

8:52 So we looked into how could we sort of lighten the structural components of the blade, and the structural components are the same as an aircraft wing. You have, like, a spar box, so usually these days carbon fiber spark caps, tops and bottom, and then the sort of sheer webs holding that whole box geometry together. And that geometry, of course, at the hub is circular, later on it sort of, too much, that arrowall shape changes, and and and all of these changes. So the point number one, geometry is varying along the radial location of the blade. Point number two, the blades actually not experiencing a uniform distributed load, it’s seeing something like that of a loading profile in an edge direction, in a flap direction, in a jet edge direction you have another loading profile. Again, bottom line, geometry is varying, loading conditions are varying.

9:54 So if you’re trying to optimize something like this, you’re looking at a problem where, for every radio location, you would, you would have to have a different structure, if it’s a ltis, then different optimized lce structure at each location. And U to know, if I mentioned the blade is 120 M long, you want to do it as a granular level, and think of the number of finite element mesh, sort of, yeah, number of elements, that’s becomes crazy. I know guys from DTU in Berlin were talking about doing topology optimization on wind turbine blades, and that’s very great, but you need a superc computer, and you probably need a month or two months of computational time on that to to solve that type of problem, so I didn’t have a luxury of that, unfortunately.

10:43 So I sort of went and looked into how else could we approach this problem, how could we optimize for every location along the blade, but sort of don’t break the bank at the same time. And we build a, like, basically the artificial neural network model that was trained on a sort of liest library, which helped with this inverse design problem. So we have had library of of latices, we trained our model, and then we said, here are the loading conditions for this blade, here is the sort of geometrical constraints, predict, or sort of give me your best solution for each location.

11:29 And that really really saved a lot of time, it didn’t necessarily gave us the final solution for for what L this geometry could be, but it very much, you know, offloaded a lot of computational work. And this is what I’m allowed to show you, without revealing too much, but basically we made a sort of ltis looking design for that spar box that sort of replaces it, and that yielded like a 27% weight reduction while keeping keeping the same mechanical properties.

12:01 So, of course, like final refinement of of of these lce structures were done again, like the forward design way, so we sort of did the final tweaks the old fashioned way. But, you knowns can really do things much much quicker, but just predicting things, and it it worked very well on this case. So, and I guess there’s couple of ways of doing it, the way we did it in that problem was we use a just simple model architecture like artificial neural network.

12:33 Now there’s more and more variz of graph neural networks, and the beauty of those are that the way the the neurons are positioned, they actually capture the geometrical intricacies of your of your ltis or of your structure, and basically that drives the accuracy up. So if you’re looking into something like this, can speak to me, or look look into this, because this is, I think, very interesting.

12:59 And I’m sort of at the moment continuing similar work with within one of our research facilities. So this is Fast Blade, it’s a tidal turbine blade, not wind turbine, tidal turbine blade testing facility, where you have, it’s probably the size of this building, of new lab, and then you walk inside, and there’s like this massive green pencil sharpener looking like frame, into which you mount your tidal turbine blade. You put hydraulic ramps here, here, underneath, and you basically Tred to break it, mechanical testing. And so what we do now with that is, we run digital twins, we run surate models, we try to predict how will those composite structures break, what delamination you’ll get, get, and so on so forth. Cool.

13:49 So now to the last part, sort of summarizing a bit, like where AI works and where it doesn’t, and again, this is my perspective, this is my opinion, if you disagree, come, let’s PE. So it’s, first of all, it’s not a silver bullet, it doesn’t work for everything, it’s not going to solve all your problems. I think the way to think about is that these are predictive tools, and they’re very good at predicting, but they’re NE, will never, in my opinion, replace like traditional PD solvers, FE solvers, sort of direct optimization. But I don’t think we should, yeah, we shouldn’t try to replace them, we should, sorry, we should try to, you know, if you have a bunch of simulation that you run, use them to train that AI model to mimic that behavior, and then do sort of maybe something like what I did, like try to reduce the computational time, like take a short cut with AI, then so you could later on refine it. But, you know, I would say don’t expect it to solve all your problems.

14:52 And then, of course, like as with any tool, you have limitations, right? So data availability is always key. AI is nothing without data, if you don’t have anything to train it on, then that’s it. So it’s, it it works much better for sort of repeated problems, or problems that are similar to what you solved before. Of course, we have this old causality relationship problem, that’s, those are blackbox models, so they need to be tested, validated, and whatnot. But, as I said, in some cases, like designing these structures, it it it still offers like a massive time saving.

15:34 And then, you know, we’re we’re moving more and more, like we we hear about, like, models hallucinating and producing, like, again, if you look at it as a predictive tool, then it’s not that surprising, weever forecasts also hallucinate. So so that, but but I think it’s getting better and better, if you if you get better and better training data. And then a lot of work, at least in my circle, is now being done on generalization, so how could you move, you know, if you trained one model, how could you transfer some of that knowledge to something else. Okay.

16:08 So this is my last slide, and basically these are the summary bullet points. So I would say, don’t use AI to replace direct methods, but rather use it to enhance them, and don’t look at exact solutions, but rather aim to reduce computational time. U, look for shortcuts that AI can, or ML can, can offer you. And as with any other tool, yeah, go and try it out, if you’re working on something similar, come speak to me, let’s have a chat, but sort of beware of its limitations, because AI at the moment is on that peak, you know, it’s it’s going to reach plateau, but I don’t think it’s yet now, so it’s it’s oversold. But, you know, let’s not overlook what it can actually offer. So thank you very much, that’s me. Okay.