Autodesk Opens Up About Generative Design

Thousands of Autodesk users, employees, and partners converged this week for the annual Autodesk University conference. AU 2020 is being held not in the Nevada desert, as is typical, but in individual offices on individual screens all around the globe.

I’ve never sat so close to Autodesk CEO Andrew Anagnost as he delivered his opening remarks.

“We’ve been working hard to make this powerful technology more accessible,” was one line that caught my attention. Anagnost was speaking of a technology that his company has been proselytizing for years: generative design.

We’ve had limited success with Generative Design, using it to see if it could improve the Golden Gate Bridge. We wanted to dig deeper into Anagnost’s comments. What did he mean by accessible?

Like any emerging technology, we understand generative design has a way to go before entering the mainstream of design. We caught up to Autodesk’s Scott Reese, senior vice president of manufacturing, to see if the promised “accessibility” had made it any closer to becoming a real-world design tool. Reese provided a candid look at the current state of Generative Design—and his company’s ambitious vision for its future.

The interview below has been edited for clarity and brevity.

Engineering.com: How many of your users are actually using Generative Design?

Scott Reese: Generative is in its infancy. Not enough people are using Generative. Some of it has to do with muscle memory; “Hey, this is just how we’ve always worked.”

Some of it has to do with its accessibility. Generative Design is too hard to use. We’ve got work to do to bring it to the masses and turn Generative into a true collaborative partner as you’re designing. But we’ve done a ton of research as to how to make it more accessible and I’m excited to get that going. A lot of our early days on Generative were spent proving out algorithms and getting them right.

What are you doing to make Generative Design more accessible?

There’s a lot of picks and clicks and loads and things to set up, so we have to make it easier to use and more accessible, and we’re working on both of those things. It feels far too much like a simulation tool today. The designs that we’re working on moving forward don’t feel like a simulation tool. It really feels like a partner that helps you explore a lot more iterations than you would be able to otherwise. So it’s all about usability.

We’ll start out by accelerating the setup processes by making some assumptions based on the information that you’ve given us. And if you don’t like the results based on those assumptions, you can just go in and manipulate some of the assumptions, add more assumptions.

From an accessibility perspective, we’ve messed around with the business model a little bit. Right now in Fusion 360, users can explore all day long with Generative and they only pay when they download a result and want to take that design further [this costs 100 cloud credits, equivalent to $100 USD]. Through the end of the year we’ve removed the 25 cloud credit paywall for commercial users so that they can explore designs. That’s obviously driven usage up a bit, but until it becomes easier to use set up, I don’t think we’ll see mass market adoption.

But we learn from them every time, just like your article [on the Golden Gate Bridge]. We get good data every time people use it. Additionally, we build in new capabilities and we want to see people put those capabilities to the test. One of the things we’re starting to pilot with a few users right now is to bring in additional simulation solvers. For example, you’re going to start to see things like fluid solvers. It’s not just about mechanical, it’s about the overall system’s performance. The dynamic simulation around fluids and mechanical properties mixing is pretty complex. If we can get the computer to explore even yet another level of complexity on behalf of the user, we’re going to just get to a better and better result. Again, all of that is predicated on making it easier and easier to use.

Generative fluids was announced at AU. When will it be available to users?

I’ve been at Autodesk a little over 17 years now. I came in via acquisition and we’ve acquired so much intellectual property over those years. Where Generative gets super interesting for me, and where Fusion 360 gets super interesting, is that we’re converging all of that intellectual property in an end-to-end process on top of a cloud data background. And then because we have all of those capabilities, be it fluids, be it mold simulations, be it tools cutting, whatever it is, now we can automate all of that with Generative.

And that’s the thing that gets me so excited, being able to see where this stuff is going and seeing that Autodesk has the depth of intellectual property to continue building this thing out for years to come.

Besides fluid, what are some of the other solvers on the list?

Thermodynamics is another thing that you could imagine, but ultimately think of any simulation capability. We want people to be able to use whatever simulation solvers that they rely on. Once people trust the simulation solver, they’re very reluctant to change. So the way that we’re building Generative out is that we want you to be able to plug in any simulation solver that you have, with whatever capability, be it emag or thermal or a different fluid solver, and leverage it just like it was a natural part of Fusion itself.

Will you be adding symmetry constraints to Generative Design? [This was a feature we noticed lacking in our Golden Gate Bridge experiment.]

We added some extra constraints, probably after you used it. [If a symmetry constraint was added, I wasn’t able to find it]. What I’ll tell you though is that symmetrical constraints aren’t scientifically better from a results perspective. From an aesthetic perspective, no question, there’s definitely human preference for symmetry. But from a simulation perspective, as much as it might defy some of the engineering logic, when you see those asymmetries in a result, run it through whatever simulation solver you want and it doesn’t matter.

We all have our favorite simulation solver that we’ve used through our engineering careers. That’s what we trust. That’s the reason people don’t often change simulation solvers. One of the most telling examples for me is we were working with a group of engineers down at the Jet Propulsion Lab at NASA and one of the engineers said, “I just can’t trust the results because they’re not symmetrical.”

Our engineers’ feedback was, “Run it through whatever simulation test you want and let us know what the results are.” Because if the algorithm is wrong, we need to go back and retrain the algorithm and figure out why it didn’t seek symmetry. They came back and they said, “I can’t explain it. But the simulation math that I trust says that it works.”

So it doesn’t mean that there isn’t a place for asymmetry, but not everything needs to be symmetrical. Having said that, we’ve built in symmetrical constraints because humans tend to aesthetically prefer symmetry.

It’s funny, we’ll prefer a less optimal design just because we like the way it looks.

Look at a chair. Why does a chair look the way that it does? Because that’s how we expect a chair to look and we know how to make it. But when you start to feed the absolute functional requirements for a chair to a computer, it doesn’t look like that at all.

In the past, it didn’t matter to find that perfect answer because you couldn’t make it. But as additive starts to scale, complexity is not only free—it’s actually cheaper, because it’s less material in almost all cases. Now once those manufacturing capabilities catch up, you actually do want to find the most optimal answer, not just the one that you happen to know how to manufacture.

Is there any machine learning or other artificial intelligence involved in your Generative Design algorithms?

Again, we’re in our infancy, but we do have AI and ML baked into the algorithm. There’s AI and ML baked into the algorithms around when you’re reaching the point of diminishing return [when the algorithm stops generating new designs], and the other is in exploring and helping to decipher the outcome [to choose between the many designs presented to the user].

How do I sift through and find the design that I actually want? We have an AI/ML team building out algorithms to help surface the most likely candidates based on what they believe your historical preferences are.

So it’s using my existing data to train the algorithm?

That’s right.

Is it tapping into all of your customer’s data? Is that going to some large database?

No.

Where do you see Generative Design in ten years?

Even in ten months, I think it’s going to be pretty different than it is today. But for me, Generative is all about making the computer a true collaborator in the process. This isn’t about replacing some of the work that an engineer does, but as an engineer you or I have to get that product out to manufacturing. And you and I will explore one, two, maybe ten different iterations of that design. Why not explore all of them? So I see in the future, the computer is going to be a true collaborator in any element of the design and manufacturing process. And that goes across industries too.

There were two other technologies released today at Autodesk University that illustrate that. One is the generative scheduling capability that was announced in Shotgun [Autodesk Shotgun is project management software for the entertainment industry]. You also saw the Spacemaker acquisition showing our commitment and dedication to Generative Design in the AEC space, helping urban planners to explore many iterations of what their urban plan might look like.