Prompt Engineering for Generative Design of Spaceflight Structures at NASA

RYAN MCCLELLAND: OK, I'm going to talk about Prompt Engineering for Generative Design for Spaceflight Structures at NASA. I'm Ryan McClelland, research engineer in the Instrument Systems and Technology Division at NASA Goddard Space Flight Center. And I'm going to talk about how you can quickly come up with these kind of organic, almost alien spaceship-looking, very lightweight designs you see on the right and left.

So the agenda for this talk, I'm going to talk about this work of Generative Design at NASA Goddard, and then give an application example for the EXCITE Tip/Tilt bracket, which some of you might recognize, and then give some more application examples, especially of some complex structures.

I'm going to talk about how Generative Design works, and then go into the prompt engineering mindset for Generative Design, and what that means, and how you get the inputs right so you get the outputs right. I'm going to talk about managing uncertainty, what you do when you don't know the requirements, and then give a summary.

So for context, this work was developed on NASA Goddard internal research funding with a goal of creating and infusing a broadly applicable process to rapidly develop lightweight spacecraft structures. And the method has been to build and test a lot of parts for a diverse set of NASA applications.

The status of this work, we can now make a typical small to mid-sized metallic structure in a mostly automated way. We can go from requirements to having a part ready for fabrication in just one or two days if it's a fairly simple part. And we've been able to demonstrate this by test.

With this presentation, I'd like to share lessons learned about the practical implementation of Generative Design and digital manufacturing, enabled your projects to improve mass, stiffness, and strength by a factor of 2 to 4, while also reducing development costs by a factor of 10. And I really do mean some massive gains, taking something that maybe usually takes a month, having it done in a few days, having a structure that weighs 50 kilograms and having it weigh 20 kilograms, and also engage this Generative Design community.

So the inspiration for this work came over the pandemic. I was looking at a lot of science fiction, watching shows like "The Expanse," and I'm really inspired by all these huge, complex structures in space that are hopefully in our future. But I'm also working for NASA, so I know how the sausage is made. And I know that these kind of things take a lot of time and are very expensive.

The International Space Station is about $100 billion project. So it got me thinking of how AI can be used in my domain expertise, which is mechanical structures, to really reduce the cost and improve the performance of space systems. And I was inspired by this quote from the Google CEO, that "AI is one of the most important things humanity is working on. It is more profound than, I dunno, electricity or fire."

So what I came across with was Generative Design, which was kind of novel at the time, but at this point, just about everybody knows what this is. You start with a text prompt, in this case, the armchair in the shape of an avocado, and then the AI gives multiple design outputs based on your input prompt or input constraints. And then the human designer collaborates to get the ideal design.

And you can see, these AI images are really kind of impressive. And some of them even have something that borders on creativity, like using the seed as a pillow. And there are a lot of applications for this, as everybody knows now, visual art, text for large-language models, music, architecture, and even structural design.

So what do you do if you need something more complex, like an astronaut in an avocado chair on the surface of the moon? Well, then you have to give a more complicated prompt. And that's what I'm going to be talking about in this presentation.

So for instance, to get these images, I had to give a more complex prompt. For one example, one of the prompts is an astronaut seated in an armchair in the shape of an avocado on the surface of the moon during the Apollo moon landings, by Salvador Dali, blue Earth-rise in the background. So I'm going to talk about how to give complex prompts for structures.

So the evolved structures process, you can think of it as prompt engineering for structures. It isn't really the same kind of generative AI that's used on these images and large-language models. It actually is just using the laws of physics. But it's functionally very much the same. So you create an input to digitally encode the structural requirements into the software-- and in my case, I've written this Evolved Structures Guide-- and then use Generative Design to evolve the optimal design that meets the requirements. In this case, I'm using Fusion 360 Generative Design.

And then you fabricate the parts directly from the Generative Design output using digital manufacturing, such as automated CNC and additive. So you can see, it can be a very quick process for a fairly simple bracket, maybe an hour to encode the requirements, an hour or so to evolve the design, and then a day to a few weeks to get the parts made.

So I'm going to go through this example in detail to show how it works. This is the EXCITE Tip/Tilt bracket. So this is a mission that analyzes the atmospheres of exoplanets from a balloon-borne atmosphere, a balloon-borne platform. And the goal is to mount this Tip/Tilt mirror assembly that you see on the right to the back of the telescope-- so pretty simple bracket, but it has some challenging requirements.

There's a set of interfaces. You have a bolt pattern on the stage and a bolt pattern on the telescope. And, of course, you have to avoid the light path and avoid interfering with the assembly itself. There are some known loads from the program office, 10 g and 3 g. And then also, we know the mass of this bracket. It's 1.35 kilograms.

The bracket has to be very stiff. The modes have to be over 100 hertz for it to not couple in with a cryocooler that's nearby. And since there's limited mass that can be lifted, the bracket has to be light, about 0.2 kilograms.

So first, a human design was made with myself and a very senior designer, someone that actually worked on the initial versions of the James Webb Space Telescope. And we took a couple days and iterated four times, which is pretty fast to do design and analysis.

So the first iteration was way too heavy. The second iteration, we just pocketed it out to make it lighter, sort of naively. And it was not nearly stiff enough, as indicated by the red. And then we got a little bit better, but still couldn't meet the requirements. These weren't very manufacturable. And then Drew finally came up with kind of a radical design that actually did meet the requirements, but wasn't really manufacturable from either CNC machining or printing. And that took us about two days. If we had another couple weeks, we could probably come up with something that met all the requirements.

Now I want to contrast that with how the problem was solved using Fusion 360. So we started out with the preserves, the green things here that have to be kept in the design, where the bolts clamp. And then we have the obstacles, the red parts where you can't have material. And then we have some other inputs, like how it's going to be machined, what the loads are, where it's constrained. And then after that, you have it evolve optimal structure. So it did these four with different machining constraints all simultaneously.

Now let's look at how they compare to the human design. So the AI Generative Design algorithm in one hour did 31 iterations of both of these. And you can see they're exactly on the mass target. But notably, they are much, much stiffer. So the stiffer it is, the less likely it is to couple in with the vibration of the rocket. So it's a big advantage for us.

So the first mode is much higher, say, about three times higher. And then the maximum stress is much lower for the additive version, almost 10 times lower max stress. So that means it's a stronger part. And they also are very easy to fabricate. This was done by automated CNC for about $1,000 in three days. And this one was made with laser powder-bed fusion. And then we made the parts and tested them, and they performed as expected.

So that was just a simple example, but since then, we've done about 40 different applications. And this is a few examples. And you can see we've done fairly large parts. This is a large 3D-printed titanium part. We've done some plastic parts. And then we've, notably, on the lower right here done some very complex parts. And I'll talk a little bit about how we're able to achieve that complexity quickly. And I really like this quote. The idea-- the best thing is not to sit around in glorious isolation and try to think big thoughts. "The trick is to get more parts on the table." So we tried to make a lot of parts.

And the Generative Design is really a paradigm shift in what the engineer does. So in the current process, you get some requirements and goals from a systems engineer or something. And a designer creates a CAD model, maybe takes a few weeks, and then throws it over the fence to a finite element analyst. And they might take a few weeks. And they might give them some feedback if it's not meeting requirements. They got to thicken this and that, but no real optimization.

Then designer makes a drawing, throws it over the fence to a machinist. And the machinist most likely looks at it, sort of curses the engineer for not knowing what they're doing, coming up with a design that's hard to machine. And most of the time, they're not going to give him any feedback other than a high cost and a long lead time.

So all these people might be in different departments. They go on vacation. They get sick. They take training. They may be very introverted, and going back and forth between them can take months or years. Now contrast that with the evolved structures process where you have design requirements and goals, but the AI system generates the design, checks the requirements, and checks that it can be fabricated all in hours or days.

And it really refocuses the engineer on requirements and getting the requirements right. So if there's any iteration, it tends to be between the requirements and the design. And we've demonstrated that if we have requirements that are known, we can go from knowing what the requirements are to having a part in our hands in less than two weeks and having that part be two to four times lighter than what a human designer came up with. And here you can see the evolution process as it comes up with the optimal design.

So expanding how that block works, obviously, it's not magic. There's a lot going on in that sort of automated block. First, it starts with the user inputs. So you digitally encode the design requirements, interfaces, loads, frequencies, factors of safety, mass target, materials, fabrication methods. And then it creates a voxel mesh of the design space.

So a voxel is a 3D version of a pixel, kind of like a Minecraft block, if you've ever played that. And it can represent any geometry at any level of detail, though, with Fusion, you're limited by the total number of voxels that you can have. And I'll talk a lot about how to maximize your use of the voxels you're given.

Then it runs a topology optimization algorithm. And basically, it's a lot of complex math in there, but it eliminates the voxels that aren't doing much, that have low stress. And it also checks they can be removed with the fabrication method that you've chosen. It checks it against requirements, and it keeps iterating until the design is optimized. The older method for this that some programs still use is called SIMP, and the newer method is called a level set.

And then the output results are made available for the user to view. You can see all what the different options are, and there's a recommendation engine. And you can see what they're like when they're reconstructed. So reconstructed means taken from the voxel space to a real CAD model.

And then the reconstruction takes place, so you can get that CAD model out. It turns the voxel mesh into T-spline surfaces. It merges the preserves. The preserves are the things, the green parts, that you need to keep in. And it cuts the obstacles, the red parts that you can't have in there. And then the user can manually tune the design, and then it's ready to make. And there's what it looks like when it's finished.

So I'm going to talk about the prompt engineering mindset, just thinking about this in terms of how to get the inputs right, get the requirements tuned in a certain way, that what you get out is manufacturable CAD. That's really the goal. Go right from requirements to manufacturable CAD with as little, few steps in between.

And getting the inputs right first every time is the goal, and it's very aspirational. You're almost always going to have to go and massage the inputs. Right? But as you get better at this, the less you have to do that. And also use computational design thinking-- so every dimension has a reason. There's nothing that you're just making up. And you'll see what I mean as I get through this. Everything that preserves, everything in the obstacles, has a reason for being what it is.

It also helps you comply with internal standards, so your internal standards can get encoded into the design. You make sure it's compatible with the manufacturing process, inspectable, and assemblable. These are all things that go into that prompt engineering mindset.

So what I recommend is creating a Generative Design prompt guide or input guide for your organization. And it helps get the Generative Design inputs right the first time, and also helps you scale adoption across your organization. And it allows everyone to get better together.

So at NASA, most of our parts are bolted together. So getting the bolted interfaces right is really important. So the bolted preserve looks like it does here in the lower right. And every part of that is determined by the bolt size that you're using. So the width of this preserve is set by the fastener diameter. We have an internal standard here where the fastener can't be too close to the edge, or you have to do some extra analysis. So we try to keep within that. So these preserves are three times the nominal diameter of the fastener.

And the grip length or the thickness of this preserve is also determined. We use about one times the diameter of the fastener, sometimes a little less. But the design of this preserve can really affect the strength of the part because that is where all the high stress is. And this will be also a recurring theme. All the high stress is where the load goes out of the bolts, so getting that preserve right is really important.

The edges of this are rounded, and the reason the edges are rounded is so it meshes well with the organic geometry that's going to grow around it. So in the upper right here, you can see how this radius here allows this to flow nicely. And then in the lower center here, you can see this photograph of the Tip/Tilt EXCITE bracket, and how nice that sort of interface looks. And there's just enough room for the fastener before the radius starts.

And the hole diameter is the last part of it. And that just used standards for-- you know, if it's a tapped hole, just the tap drill size. If it's a clearance hole, then you can use something like the ASME standard. And I mentioned the radii.

So bolt interfaces also have obstacles. And again, those obstacles are predetermined by the size of the fastener. So the obstacle diameter is the washer or fastener diameter that you need to get on the preserve to clamp it, plus any hole tolerance, plus two times this radius that you put on the bottom of the obstacle.

And the reason you put that radius on the bottom of the obstacle is so when the material grows around it, like you see here, you get this nice radius in the bottom of your part, which makes your part stronger. So you can see in this image of these speckled pieces that that is where the highest stress is. So these parts failed. They were plenty strong, but we tested them all the way until they broke. And you can see they failed right where that radius is. So having a radius there is important, and you get that radius by putting a radius on your obstacle.

The length of the obstacle is driven by the fastener length. So if your fastener is long, you need a longer obstacle to get it in there. And then you also have to consider any kind of tool access that you might need. So my recommendation here is that you use Configurations, which is a feature in many CAD programs and a new feature in Fusion 360, that you can use Configurations and have these already set up. So you have a Configuration for whatever your standard bolt sizes are. You know, if you have a clearance hole, a tapped hole, a hole with an insert, you can have Configurations and make it very easy to create this geometry quickly.

And you can also use the Obstacle tool that's built into Fusion. And this is a more advanced obstacle where the fastener might need to come in from the side, if material really wants to grow around it.

And you also need Obstacles to avoid other components in the assembly. So a couple recommendations here, you want to simplify that as much as possible. So if you look at the Tip/Tilt assembly here, I've just simplified it into this red box. And that allows faster modeling. You don't have to have as large of an assembly. And another thing it does is it prevents voxels from showing up. So some voxels would actually show up in this empty space, potentially, and use up some of your voxels that you know you don't want there and they're going to get removed, so really simplify your obstacles.

And then, also add radii on your obstacles because, again, that obstacle will be cut. And as you can see on the rightmost image where the obstacle sort of cut itself out of the part, there is a radius so that that doesn't cause a stress concentration and can be easily made.

And then you can also add Obstacle Offsets. And this is something that a lot of people may not know. The obstacles are cut from the final design with, like, a-- you can think of it like a Boolean operation. But the offsets aren't cut. They just sort of discourage material growth. So you can also use Offset just to push the material away from your mating part a little bit.

So I recommend you create preserves and obstacles in your native CAD software and keep them parametric. So you might use a different software for most of your work and then import things into Fusion for Generative, which is what we primarily do. And if you make the preserves and obstacles in your native CAD, and then things shift around, they will shift around parametrically. So it makes it easy to export them and re-import them into Generative Design.

So an interesting thing we do with loads is we use remote mass with accelerations a lot of times rather than forces and moments. So for our applications, our structures are usually supporting a bunch of mounted components, optical assemblies, detectors, electronics, tanks, communications equipment. And I recommend representing mounted components with Remote Masses.

And it's a very fast and simple two-element representation of another part that's mounting to you. So in NASTRAN terms, it's a CONM2 with an RBE3, and it doesn't make any assumptions about the stiffness of what's mounted. So it's conservative in that way. So if you have a very stiff electronics box that's going to actually add stiffness to your design, that's not represented. So your part will be even stiffer than you had anticipated when you finally make it and integrate it all together.

And then we apply acceleration loading. And you have to apply the acceleration loading in three separate load cases. It's very important not to just have one load case with three different vectors because those vectors just get summed. So you need three different load cases, x, y, and z loads. At least, that's the way we do it.

And then we have this mass acceleration curve, so we know what acceleration to use to get started when we do a design. Lighter components see more acceleration. Heavier components typically see less. And then on the bottom, you can see we actually use this mass dummy to represent the assembly when we tested it because we didn't have the assembly on hand. And that mass dummy, basically, in the analysis or in the Generative Design is represented by this mass element.

OK, so now I'm going to go to the demo. This is the EXCITE Evolved Optical Bench, a very complex, fairly large part that we made. And here you can see my end customer, very happy to have his part. OK, so this is the CAD of the final part. And this is what was used to input, to create that part. So here you can see a bolt preserve. It has the features that I was talking about, the right size, the right obstacles with the radii. And then that turns into having a really nice-looking bolted interface.

And some other features, we had a more advanced obstacle here because we were getting material growth in places that we didn't want it. So we had to extend that bolt obstacle out. And then you can see these bolt obstacles here are rather long, and that is because they needed to preserve room to get fasteners in. So you can see in this design, fasteners needed to come in from the side here. And we ended up having this obstacle cut there.

If we look at some of the other obstacles, here is the obstacle actually representing another optical assembly. And you can see over here where that obstacle ended up being cut from the design. And we put a nice radius on it, so you get a nice radius down there.

So yeah, I think this part came out really nice. It was fabricated, everything worked, and we had to make very, very minimal changes between the output of the Generative Design and the final part as we fabricated it. So yeah, it's a pretty complicated part, but just a bunch of really simple geometry, rounded rectangles and cylinders and stuff. But the part that came out of this was, obviously, very complex and very optimized.

OK, so next I'm going to talk about design objectives. So we generally prefer max stiffness with a fixed mass target. NASA designs are usually very stiffness-driven. So we want to take what we think the mass could be and then see how stiff of a design we can get, again, that avoids the vibrations of the rockets at lower frequencies.

And we're usually not pushing the envelope on failure by using something like minimized mass. And also, with the minimized mass, we often find that the maximum stress is at the interfaces in these designs. So in the upper right, you see one of these EXCITE brackets. And the maximum stress from Generative Design is about 5 megapascals.

But then when we take that into simulation with the same load cases, same bracket, the max stress is 11 megapascals. And that's because the voxels don't capture corner stresses as well as the tetrahedral elements. It's just a little bit different in the way the analysis is. So usually, we don't want to drive to max stress. And that is also actually the reason we usually keep a factor of safety of two in the simulation.

So you can also use Modal/Buckling/Displacement additional constraints if you need. I've never seen a design yet that buckles, but these other ones are very useful, or if you have really thin things. And in general, I'd recommend using max stiffness starting with the mass target a third of the human design if you had one. So for whatever reason, if you take a human design and a Generative Design, the Generative Design just ends up about a third of the mass. I'm not sure why that is, but I've just found that across a lot of different applications. And I find that most customers, if you cut the mass in half or better, they're generally happy and don't need to worry about eking out the last 5%.

And the other thing I recommend doing is to vary the mass target up and down about 10% or 20% and just see what looks right. So for example, in the bottom image, the image on the right just looked more right to me than the one on the left. So this is where some of the human element comes into play.

Now I'm going to talk a bit about fabrication, CNC versus additive. So a lot of people see these parts, and they immediately think, oh, you have to additively manufacture them. And that really isn't true at all. Organic structures can generally be CNC machined. And state of the art CNC is far beyond what most engineers expect.

So in the middle image here, you see this kind of wild-looking spaghetti of metal that was CNC-machined from a single block of aluminum. And the advantage of this, or one of the things that was kind of surprising in this research, is that because of the way the optimization algorithm works, these designs are actually pretty stiff and don't deflect when machined. So it was kind of a surprise that a lot of the machine shops told us, oh, we didn't expect this to be as easy to fabricate as it was.

Now AM has some real advantages. You can get improvements in stiffness due to more design freedom. So you can have parts that, like in the upper right, that have sort of hollowish areas. And you can also make big titanium parts, make parts out of more expensive materials sometimes cheaper. So in the lower right, you see a titanium 3D print.

And Additive Manufacturing, AM, generally has cost and schedule disadvantages for aerospace. There's a limited source of AM vendors now, especially for large parts. You know, I had to go overseas to get a large part made. Material properties are process-dependent. So you can take the same design, send it to different vendors, and they could have different strengths because of the way that the laser traces or whatever.

The tolerances are relatively poor compared to CNC, so a lot of times you have to machine them anyways. Surface finish is rougher. You have to remove supports, which requires manual finishing, which can lead to variable quality in the surface. They usually have to be heat treated. And for NASA, there's a bunch of additional testing inspection required.

And something that surprised me is when you send one of these off, often, they'll tell you, well, it might fail the first time to build. And we might have to build it a couple times to get it right, which is usually something you don't encounter with CNC machining.

But in the future, there could be a lot of advantages to additive. So my recommendation, really whether it's CNC or additive, is to talk to the people that are going to make your part after you do a first iteration. So do an iteration of your design, and then talk to, ideally, the CNC programmer that's going to program the machine or the additive manufacturer's slicer operator. You want to talk to the people that are really making your part and get input. That has been really powerful.

And I also recommend running an unrestricted fabrication option when you do your Generative Design to see what the ideal design might look like. So maybe you end up with a design that, oh, that actually only needs to be machined from two sides. So it can really help inform what you choose for manufacturing when you see that unrestricted design.

So on the two parts shown on this page, these were both made from only two sides. They were machined in only two setups from the top and bottom. And you can get a lot of complexity from just machining from two sides. And I've generally found that when you select the milling, Fusion checks that you can remove the voxels, as it removes them with the tools that you've selected. From the axes you've selected, those tools can come in. And so far, that's been accurate. So when I've sent something out to be machined with a certain set of machining constraints, it's always come back machinable.

And the fewer tool directions you have-- so you can choose either five-axis or three-axis. And in three-axis, you can choose any of six directions. The fewer directions you have, the easier it is to machine, reduce number of setups, generally, try to minimize that if you can.

The minimum tool diameters I choose-- you know, if you have a larger tool, it's generally going to be faster to machine. And then a smaller tool might take longer. I generally use about 1/8 of an inch diameter tool. The shoulder length, so you know how long your tool is, determines how far it can get into your part to remove material. For aluminum, 10 times the tool diameter is what some machinists told me to use. And there's also this head diameter, and that is what the tool is mounted to-- again, determines how far you can get in to reach.

So it's possible to do some very complex structures. So you see on the right some really absolutely wild-looking complex structures that are possible. And the complexity is really driven by the number of interfaces. If you just have two or three interfaces, it would only take a day to design, analyze your part, and prepare it for fab.

But the development time increases as you add interfaces, like, maybe even to the square. So if you have twice as many interfaces, it might take four times longer to get it right. But very complex structures are possible. And it's great for consolidating parts, so this design on the upper right consolidated a whole bunch of parts into one.

And it's also great for mounting things that are in weird configurations. Like, in the lower right, you have these planes facing all different directions. And that would be very hard for a human designer to accommodate, so really powerful for making complex things.

And as you do this, though, if you have a lot of interfaces, beware of lightly loaded preserves or components. So if you have a structure that supports 5 kilograms, and then also supports something small that's only half a kilogram, you could end up with failures to process, converge, or weird results.

So make sure, first of all, all your preserves have to be either constrained or have some kind of load on them. And if the load is too light, it can cause problems. And the reason that is is if you have a highly loaded part, and it's removing the material that isn't as highly loaded, but then you have just a little bit of load over here, it'll actually remove the material that's supporting that part.

And you can watch out-- sometimes it becomes disconnected, but there's been some changes now where it actually will try to connect it. But you might get some weird-looking stuff. So I generally find you either want to increase the smaller load to be like one fifth of the largest load. You might have to play with that a little bit. Or I think you might be able to just put the smaller load in a separate load case, and it'll make sure it still gets connected.

So also, with these complicated assemblies, you have to really think about how all the components are going to be installed. So in the upper right, you can imagine if something was coming down from the top in this, it might hit this crossbar. So you might have to add some additional obstacles to make sure something can be installed.

And another problem that you run into with these large, complex systems is running out of voxels. You could run up against the voxel limit. And if you don't have enough voxels to represent your preserves-- you know, get a couple voxels in your preserve like these flat parts here-- you could have problems with the study. So I'll talk next a little bit about how to manage the voxels.

So managing your voxels-- so hopefully, you've played a little bit with the Study settings. The Resolution, whether it's low or high, that determines the total number of voxels you have available to you when you start iterating on this. And the more voxels you have, the slower the runtime is going to be, but the more detailed and complex things you can work on.

So I generally recommend using the lowest resolution setting that gives you a good result and that successfully reconstructs. And that means you get faster runtime, faster iteration time. But be aware that the part performance can be sometimes a little better if you use a higher resolution setting.

So on the right here, you see a low voxel, or, like, a coarse design with a low setting. And then you see a fine design with the high setting. These are the exact same study, but you can see clearly there's a difference between the outputs. Coarse generally gives you less complex designs that are sometimes easier to fabricate. Fine gives you more complex designs that are sometimes have better performance, but it's usually not by a huge margin. And, of course, there are aesthetic differences to take into consideration.

So my recommendation here, before you fabricate, try the highest resolution just to check if you get any improvements in performance and aesthetics. So it's just another thing you might want to vary, like you're varying the target mass a little bit to really get it fine-tuned to what looks good.

And there's another trick you can use to get some voxels back. And that is when you have, again, these very complex models, you can become voxel-limited. And small preserves might not be represented well. So a recommendation here is you can add obstacles into empty spaces.

So if you look at-- this is the same bracket. This is iteration 1 out of 34 iterations. In every solution I've ever seen, there's no material here in the end. And you can see that in iteration 15, all that's totally gone. So if you want to get some voxels back and maybe increase your runtimes or allowed you to have a more complex part, you can put an obstacle here because you know there's not going to be material there anyways.

So one of the most challenging things is not necessarily the Generative Design itself from a real standpoint of making parts. A lot of times, the single longest step in Generative Design in terms of calendar time is getting the requirements from the stakeholders, the people that need this part from you. The stakeholders often want to get the requirements perfect before starting the design because they're used to design and analysis being slow. So they want to get it right the first time. So you really kind of have to train folks to think differently.

And a lot of times, I find I run into this chicken and egg issue, where the design engineer is asking the systems engineer, oh, how light does this part need to be? And the systems engineer says, how light can you make it? And then you say, well, how light does it need to be? So you can really end up in this loop and not getting anything done.

And I recommend that the best way to get to the requirements is to make a part, or make at least a design and send it to folks. So you just make whatever assumptions you can about the interfaces, loads, what the mass target is. You just make a design because maybe it only takes you a day or so.

And that shows the stakeholders how fast Generative Design is to iterate. You're like, oh, I have a design. And once they see that, a lot of times they'll be like, oh, well, it's not right. It needs to-- you know, you're not considering the loads here, I can't integrate this, and now you're getting requirements. So it's a lot easier to find missing requirements after you have a design. And also, it shows that design and analysis are now fast.

And inevitably, you're going to miss requirements, especially for very complex designs. So really, you want to get a part, get a design out, get an analysis out, and it allows you to iterate really quickly on the requirements. And it's one of the things I've found has made this a powerful technique.

So in summary, Generative Design can lead to revolutionary improvements in both performance and cost schedule. So we see something like a 3X improvement in the strength, stiffness, weight of these parts. And we're getting it in 10 times-- 10X less time. And it's really a new paradigm in design engineering and kind of changes what the engineer does a bit.

And the prompt engineering mentality helps you fully realize the benefits of the technology by helping go from requirements to manufacturable CAD as quick as possible if you get the inputs right to get the outputs right. And I recommend that your organization write a guide for your applications that are specific to what your requirements are, the kind of problems that you encounter, the kind of fasteners you have, the kind of loads that you see. And once you have that written down, you can iterate on it, and you could have continuous improvement on your process.

And in summary, no more making things up in design. So these designs that we're making, they're completely computational, if you will, that there's not extraneous things going in there.

OK, that's my last slide, and I hope you enjoyed it.