0:03 Hello. Hi everybody. My name is Marco. I’m the CEO and co-founder of ToffeeX. So today I want to talk a bit more about topology optimization. I’m sure everybody here has some knowledge about topology optimization. And I want to start from my story, the story of Tofax. Everything really started around seven years ago when I left Italy. You can tell from my accent I’m I’m Italian. Where I will study maths in Italy.

0:34 I moved to London Imperial College and I met my supervisor. It was crazy around the introduction of add manufacturing into application and aerospace. I arrived to London when there was the the huge news of cement successfully 3D printing a turbine blade. So a very critical component for a gas turbine that was reliably 3D printed. So that open a huge gap in the market which means okay now we can literally I mean now in 20 years we can literally 3D print also critical components.

1:13 So we are not constrained anymore with traditional manufacturing technique like laser cutting or stamping and whatever. So why are we still doing straight channels for fluids? Now we can do crazy. We can do veins inside gastrobines. And my supervisor was very very straightforward. I’m an engineer. I don’t get the deep maths. You’re a mathematician. Add veins into topology optimization please. So it was a long work of heavy maths and we managed to crack what is topology optimization for flow dynamics.

1:48 I’m sure all of you know topology optimization for structural problem. So seeing all the structure that look like bonds no not surprisingly but structural is mostly linear. When you put together topology optimization and nonlinear problem like Nav Stokes or flu dynamics things get very messy. We did it and we found a company a few years ago. The name is toffee for the first time ever. I show the acronym because I always forgot which stands for topology optimization for fluid and energy engineering.

2:21 So it’s really we shape it into a product where the customer can start from an an initial design mostly in application for fluids and heat exchange for example a heat sink or a heat exchangers and it can create h can find what is the best shape for that component by leveraging the physics by leaging the physics simulation on around that component. So in a nutshell, how can I go from the the solution on the left which has been designed only by intuition by engineers to the solution on the right that has been carved by the physics by the fluid on the engine.

3:09 We shape it in this way. We really care about the psychology of the customer. That’s why here today I want to talk about the sense of explanability, the sense of control and the psychology of the engineer. We shape it in a way that is very familiar to pretty much all the aerospace automotive all the engineers. So you have a 3D visualizer and you have the possibility of setting the physics around the space the fluids, the velocity the manufacturing constraints, the material, the minimum channel size etc.

3:48 Once everything is ready, the software simulates the physics, the flow dynamics on that component. Then it use something an optimization technique a joint technique to literally understand step by step which region have to be solid and which region have to be fluid. The process look like this. The component let’s say the fluid it’s is carving the shape the material and creating those gray structures that are literally the final solution.

So splitting the fluid impinging the fluid saving the cold fluid for the areas where the the material is warmer. All this stuff are done automatically. Now this is the core. This is the engine. It’s not that easy to make it work. It’s not that easy to go to a customer and saying to an engineer and saying, “Okay, now you can use it to create a heat exchanger for the next generation of aircraft or for a coolant plate for a battery that is going to go on the market in three years.

5:06 Why the the software is very powerful and as soon as you change the physics is changing completely the outcome. Those are just example the real example of the outcome of the software which we have a vaporizer we have an exchanger for a fun we have a part this was made and published by by Toyota is a coolant system that managed to save a big amount of aluminum.

5:34 The point is the when you have such a powerful engine that you feed with the physics and is giving your results, it’s very easy to fall into the trap of the black box. I call it the the blackbox mindset. So you ask the software something and you’re expecting the solution to be the most efficient one on the one you really need and it’s not printable and you go back and see okay I’m going to change the algorithm in all these loops of evolution of the software something that tends to be forgotten is the role of the human now when you design something is a creative process so it’s a cognitive exercise, a cognitive pro process which pretty much involve three main areas.

6:29 One is the abstraction. Even when I do a drawing, I’m abstracting what I want to create. Then is the physical intuition. If I want to do a radiator for a car, I I understand that the fluid is going to move in a certain way. I know it before building, before drawing something. And then is the critical thinking. Am I doing the right things? Is this still working?

6:50 Now I got this. What should do next the the decision making process now historically we have been given a lot to the abstraction all the cut software all the CE software etc now in the last 5 10 years with topology optimization with surrogate models with AI we are giving a lot to physics intuition and it’s here that we are making the mistakes sometimes of falling into this blackbox mindset because when we give too much to the physics intuition.

7:22 So we let the physics design something. We tend to remove the critical thinking. So what I’m suggesting here is we need to go we need to use topology optimization in a way that you have very short iteration. You give value to the human and then is the human deciding to the next step. Now why this? Because if I want to put a heat exchanger on a car, if I want to put a heat exchanger on a new aircraft or a new car, I need to be 100% sure that this is working.

7:55 I need a sense of explanability. Now, what is explanability? It’s really something that we can even identify with beauty. Is the recognition of some patterns that I in consciously recognize as trustworth is I have a component that works better but is very close to something that I’m familiar with. Now how can I include this explanability in topology optimization that is traditionally extremely complex and the solution are organic and hard to understand.

8:29 I’m going to bring two test cases where we developed some feature to help the sense of explanability while during the design process. So the first one it’s something actually I can say now this is a use case for for Nissan. We got the the clearance to say it just in time. The numbers are slightly different. So what is this? It’s a device that exchange temperature is a oil cooler.

8:58 So you have the oil warm oil flowing in one direction and you have air flowing on the opposite direction. Now we developed the possibility of doing topology optimization for fluids using two fluids at the same time. I picked this because it’s the the amount of results you can get when there are two fluids. So two nonlinear system interacting it’s crazy. It’s like even if the solution are good, navigating the the multiple solution you can get was was too hard.

9:31 So we have the warm oil going one direction, the cold there going in the other direction. What I in the platform the the desend domain is this cube. So I want the the heat exchanger to be shaped inside this cube. And as usual I set all the boundary condition and I click start. So let’s see if I only want to optimize for pressure losses. So I want to find the best path for the two fluids to flow one around each other without exchanging any heat.

10:02 This is extremely intuitive. So the oil will go in the center and the the the coolant the fluid will go around. So you see here this is the optimized structure which is just a channel and then the red one is the shape of the oil and the blue one is the the coolant that goes around. Now as soon as I add heat exchange, so in my objective of the optimization, I want the temperature of the oil to go down and the temperature of the coolant to go up because up because it’s extracting energy.

10:38 This things start to look like this. So is extremely intricate. The fluid start branching. The big scale got branched into small scales. He got impinching inside the warm fluid. And as you can imagine, they want to increase as much as possible the surface of exchange and keeping the pressure losses as low as possible. Now, this will never end up on an aircraft because there is so much I don’t understand.

11:10 Even if I do all the testing, even if I do all the the structural response, the fatic testing, still there is a lack of explanability of the engineers and this will never be produced and put on a the next generation of Boeing of Airbus aircraft. So what we do, we want to enforce the something for more familiar so a pattern inside the solution. So what we did is modifying the solution, modifying the design to take into account a repetition of unit cell.

11:42 Some things that may look like a latis or a gyroid, but it’s not quite because we are still optimizing the full design all at the same time. We are just enforcing the solution to have a repetition. And if you look at I am forced like on the top there I enforce the solution to be a box that is made by 2x 2×2 unit cells. What is the advantage?

12:12 I still optimize the full physics. I still have the measure of the global efficiency but I gain on manufacturability. I gain on a certification. I only need to work on one module and then it’s easier to extrapolate. I can do three by three by three and I can continue on 5x 5x 5. Now here I was cheating because if you go to the unit cell is too small you need to increase the mesh resolution otherwise there is a mismatch of the scale of the heat exchange.

12:47 But the point is there. But what is really the value? Now if you look at comparison of the performance of the the monolithic solution, the one has been optimized all at the same time and then you compare to the other. What can we say? Nothing. I mean it’s nothing more or less the same performances. Yeah I stretch more surface. So obviously more it exchange. What is really the value we’re giving here?

13:16 It’s the sensitivity on the initial data. Look at that. The one on the top there is the monolytic solution. I only changed a 50% of the velocity of the oil. I got something completely different. So it’s unexplainable what’s happening behind that. Even though the solution is good. I did the same of the 3x3x3. The solution is still there. The channels are deformed to contain the new pressure.

13:41 But the solution is still there. I can explain why varying this velocity the channel get this difference. I can explain why there are some pins arising etc. Something else we have been focusing it’s been close to something that you know already. So this is an example of a component on a gas turbine is an injector. And u so traditional topology optimization you would take a a design domain and you would say okay optimize for pressure losses for mixing etc.

14:12 And this is was the result again why should I put this on a gas turbine or at least I can do in 20 years but not tomorrow it’s so different for what it was before. So what we did was okay let’s simplify let’s take at least the skeleton of the fluid there and we implemented a function that allows you to measure how far you are from your initial design and then we put it very tight and this was the first solution a de a deviation of 10% which looks very similar and then we can relax okay now I understand what’s happening I can relax a bit I can make a deviation that’s 25% and I can make a deviation that is 40% which was the one like unchained.

14:59 What was extremely surprising was this if you look at the baseline so the the the geometry not optimized that I I normalize as one the 10% so a solution I get literally in 20 minutes using toffee without checking manufacturability etc I already gain the majority of the the increase of efficiency all the rest it’s still additive is still good but it may balance out with all the you know the studies around the manufacturability etc etc.

15:38 So here the value is if you’re close enough to something you know you can already h get a big push an increase of efficiency at least to start at least is an easy start and then pushing with more intricate organic design so that sense of control could be already enough what’s next and then I’m going to conclude putting together both. So this is very beta version hopefully will be for for Barcelona and doing the modularized structure but relaxed.

16:12 So the modules are pretty much all the same but slightly different so they can capture the local efficiency the local gradients of temperature of the loading of the stress etc etc. You can tell here even though it might not look it may look all the same if you look at the the holes there the channels path they’re slightly different or about 10% which is a lot of units that are only two position apart.

16:39 So just to recap in a nutshell effective generative design implies high controllability. I cannot think of generative design as a black box a magic black box that just cut the human outside. So and to do that we need to control obviously the physics the meshing the resolution for the 3D printers etc. But we need to control the design process. So include the possibility of recognizing and controlling patterns capturing the physics with the scale I want with the position with the shape of my patterns and the unit cell and remaining at least on an initial phase as close as possible to design for which I’m familiar with. Thank you very much.