0:01 Up, all right, let’s do it, cool. Hi everyone, I’m Joe, this is Laurence, we’re co-founders of Generative Engineering. We have backgrounds in various different industries, Lawrence is more the academic computational design work with MIT, Stanford and Cambridge, I’m the exact opposite, the university I graduated from was possibly one of the worst in the world, but I’ve been building some commercial products for a few years now, and we met building electric vehicles, and that’s when we realized what a lot of other people realized: the engineering is complex, engineering is slow, and engineering is ridiculously expensive, leading to many numerous failures, including the company we worked with as well when we discovered all of this.

0:51 So we wanted to try and change that narrative. We wanted to build a solution that was a general solution, that didn’t look at specifics within the engineering process, but really wanted to try and tackle the fundamental issue that we saw impacting every single bit of engineering, which is really looking at how engineering struggles with traditional projects, but also struggles with projects looking at new technology as well.

1:18 So here we have a number of examples of just how many many different engineering projects fa, and usually comes down to the same fundamental reason. So I wanted to try and give you an understanding of why we’re doing what we’re doing, and Lawrence is going to fill you in on the actually what it is.

1:35 And what we realized is efficient decision- making is usually the blocker for every single thing when it comes to engineering. The complexity of engineering, the cost of engineering, comes down to everyone’s abilities to make efficient decisions, which is ridiculously hard and challenging to do. Hard trade-off decisions absolutely everywhere, and the mechanism for making these decisions, not really fit for purpose. Design reviews, we kind of see, is just a load of engineers angrily shouting at each other, and the process that it involves is not one that has any real decent outcomes.

2:13 So we wanted to try and see what we could do if every engineer could make efficient decisions. Can we give engineers all the knowledge they usually gain at the very end of an engineering process, at the very beginning? What if we could enable the existing engineering stack to work very well culturally with a lot of engineers as well? And what if everyone else who works with engineers could understand why these decisions are being made as well? Will that bring world peace? Will that not? Will that mean we can actually make some better products and enable a sustainable planet, which is very much what’s at our heart. So that’s framing why we’re doing what we’re doing, and Lawrence will tell us exactly what.

2:57 Thank you, Joe. So, as you might have guessed by a company name, what we’re doing involves using generative models to make data driven decisions. So, like Jo’s been saying, we want to enable people working in hard technologies to know whether they should be using additude manufacturing or not, what is the trade-off between cost and efficiency, get the data to make these decisions upfront.

3:17 But like other people have done, we need to define things, because generative can mean anything to different people. So the way we see it is, well, firstly, we could ask a generative model what a generative model is, and we get some boring answer about how it works underneath, and maths and things like this. From our point of view, what we want to get to is enable people to do things like this, where if we start with something like the boundary conditions and a domain, we can do topology optimization and we get to a geometry, this is generative.

3:45 We could, if we’re using some sort of AI system, say, design me a pump, and hope to get out the other end a design that would be really generative, or maybe some requirements, and we plug this into optimization, or exploration, of, or design of experiments, we get lots of designs. So we want to enable people who are already doing things like this, which is quite a few people in this room, to take this into the messy complex world of engineering and make decisions with it.

4:09 But firstly, you know, spitting out one design is interesting, what might be more interesting is if we get a design plus, or the specification for it, so we know maybe how it’s going to perform under different simulation results, dimensions and things like this. But, as the guys from tof X have already explained, like one design, when it comes to making decisions in a complex messy system, isn’t usually good enough.

4:30 So what we’re trying to do is let you go from these high level inputs to a trade-off set, or a number of different directions, or a way to communicate: if we went in this direction it would mean this, if we want the cheapest design it would mean this, if we want the best design, high perform design, it would mean this. Because then you can have all the stakeholders and all the customers saying, actually, this is why we can’t go in this direction, rather than having to do manual iteration to get back to that answer.

4:56 And the way that we want to do this is by us, us the tools that are already in the industry, like we’re not really in the business of telling you how you should be creating your generative models, we just are saying this is the way the industry is going. Like we’re in this room because people, I think, agree with this, but the world of manually iteration and manually drawing boxes on screens forever is slowly dying, we hope. But I’m putting these on screen because they’re what we use for the examples I’m going to show in a second, and because they’re startups with apis that actually are collaborating and building part of an ecosystem, rather than than some of the more legacy engineering tools that make it specifically try and make it hard to integrate with.

5:35 So to illustrate this let’s look at a pump. So I’ve been showing a pump on screen, I think it’s an example that’s been on this stage a few times, it’s a classic engineering problem that has been solved lots. But in this case what we do is let you, in a tool like on shape, define a parametric C model as you normally would, and define the parameters that would define the way to create the geometry, and in Sim Scale create some templates like you normally would.

5:57 In this case we use one to do CFD, we do FAA on the stress, and we do a vibration analysis, and all, all, all that we’re claiming that if you have some software magic on the back end that just plugs them together, you can have a generative model, and you can start to generate data to make these decisions. You may have noticed that that gift is actually from OV, because it’s a better way of doing this, but we think it’s important to actually bring this to people working in b as well, because there’s no reason why not, it’s just not as good, so we hope to move people away, but it can still work.

6:29 So now let’s dive into our software. So if you have defined this in on shape and Sim Scale, then you move, move over into our tool. Behind the scenes we’ve done some glue magic to connect these things together, but then we can ask, design me high efficiency pump, and what this is going to do is not try and be an engineer itself, it’s just an interface to the complex system we have underneath. So it’s going to look at the parameters that are in the model and in the simulation, and say, if you care about efficiency, you should look at this output called efficiency, and these other geometric inputs, which, given the data I have, are correlated to it.

7:02 So the title of this talk was something like the gap between AI in knowledge, and in engineering, and for us AI at the moment is mainly a really good way of getting the barrier of doing these complex things lower. It’s not yet, especially because if you’re working on new problems generally data doesn’t exist, it’s not yet a way of actually doing any of the engineering inside.

7:23 So then we can ask as well, like, okay, well, if I care about efficiency, what are the trade-offs, because we don’t just want to get to a single optimal design, we want to get to something where we know if we have the most efficient, what are we going to lose. And in this case, it’s again using the data in the system and all the information that’s already available, to say you should probably consider these things. But, like I said, like we’re treating AI here just like an interface addition, it’s not really meant to be replacing anything, so if you want to go underneath and say, actually, manually let me choose these parameters and let me set up the problem how I want to, you can do that as well.

7:54 And so what this does is essentially turns what we had before on the parametric C in Sim Scale into what we would say a generative model, because I can now work at a high level, start to generate data and then take that to decisions. If you want to discuss whether that is a generative model or not, then come have some magic mushroom tea with me, but we claim that it is.

8:11 Then what you do, well, move over from experiment into discover. So fast forward a couple of hours, we run this through the optimization and the exploration algorithms underneath, if we had no objectives we’d be doing design of experiments, and we get this thing that just lets us very efficiently get towards insight.

8:32 So here I’m going to say, well, I was interested in efficiency and stress, and now I can see all the generated designs, not just a single one, and for each one this specification of all the output. And the system underneath has handled all the things you don’t really want to, as an engineer, like where do I put the cad, how do I pull out the visualization, how do I show that in a browser, how do I do that for a 100 designs, how do I do that for 100 designs that are changing over time as I change what’s in my geometry model. This just tries to take as much of that away as possible and let you get to data.

9:00 So in this clip here I’m just going to make a few different visualizations, and then just show you, like, how easy it is to handle all these different simulation results and all these different geometries. But the key thing we want to do here, is, like I said, is not just stare at a data set, it’s get to a screen like this, because this is now saying, okay, well, if we want the highest efficiency design, is that one on the right, if we want something a bit cheaper, it’s the one on the left, and maybe the one in the middle is a tradeoff and it has an interesting behavior under natural frequency.

9:30 And we want to enable engineers who are using generative models to take this to their team, because this is now completely sharable, and this is running on a web browser in the cloud, we don’t need to like manually do anything locally, we just say, hey, let’s look, let’s figure out what we want to do next.

9:44 But again, the title of the talk is a little bit about AI, so if you really wanted, to llms as a brainstorming partner, when you have this much data, is an efficient use of AI, because it is interesting, but a lot of manual work to go through 100 explored des designs, looking at all the insight, all the generated results, and saying what is in here. I mean, it can be really useful, but it’s another opportunity for something like AI that has access to the raw data set underneath to give you suggestions.

10:12 So here, very quickly, I can say what is in this data, and it says, look at this really strong correlation, and here’s a design that could be useful to take to a design review, very quickly. And just to reemphasize, so I go back to this screen for a second, because what this AI is suggesting to you is, here’s a correlation, an actual number, I’m just pulling that out of the data for you, here’s a a number for performance, pulling that out for you, just saving manual work. It’s not really trying to say here’s what you should do, which we think is like the way that II should be used at the moment.

10:41 So this was a not that interesting example, we’re at CDFAM, we’re not that interested in on shape and Sim Scale, even though Sim Scale is a really really good tool and we love it. So let’s look at an example that might be a bit more interesting, because Joe explained we think this is especially useful when you’re working with new technologies that have an even larger design space than something like traditional CAD.

11:01 So if we’re looking at this wheel, which is actually a wheel to go on the space mobile power station we showed the generating on the first slide, we probably need to consider even more data. So here we create a generative model in Python rather than on shape, it can spit out a wide, a variety of wheels, and we’re going to consider the implications of this.

11:23 So let’s skip to the good part, and we’ve gone through the stuff I’ve just been showing about the pump, and we very quickly see, hey, look, massive trade-off between the rolling resistance and the mass, which we might have known already. But there’s some really interesting, like, sharp corners in here, and there therefore some interesting stuff to take to a design review, because if we’re sending this thing into space, mass is obviously important, but if it’s a vehicle, rolling resistance and range are going to be a big performance indicators.

11:46 So, within a few clicks I get to a screen where I’m saying, look at these three things, let’s discuss the trade-offs, and then make a decision. And the thing I want to end on with this example of the the product, is because this is all integrated and sharable, like, anyone can go back from the data we’ve got in discover to go and set up a new generation.

12:03 So on here I can say, from, from the previous data, I saw that mass was going to be really much of an issue across that sharp knee point at about 0.4, so let’s adjust our requirements, set a constraint on that, and then instantly can go click generate again and get to new data, and we can get into an iterative loop of making decisions, generating data, making decisions.

12:24 So this is interesting, but the thing I want to leave you with is, we’re a start, we’re very early, like, we think that this is solving some of the problems we’ve seen in industry before. But the reason we’re doing this is because we want to build software for a future where generating designs is the norm rather than the exception, and I think a lot of people in this room would probably agree.

12:42 You know, Brad, yesterday, from enop, gave really good motivation for why manually iterating is the problem, generating is better. But then what does engineering, as a collaborative messy human centric process, look like when you can generate designs as easily as that? That’s what we’re building towards.

12:57 So if you have opinions on what that should look like, whether what we’ve done is a tool, the right thing to do, come and chat to us. If you want to know more, here’s our website, email us, and you can get early access to the product if you want. Thank you.