AI-Powered Design: Expert Takes

As a little experiment, this DE editor turned to a popular AI-driven chatbot tool to see how it would define AI-powered design, and whether that definition, delivered in a millisecond, lines up with the experience-backed perspectives of several vendors in this space.

Here’s what was learned:

• AI-Powered Design as defined by ChatGPT:

a type of design that uses AI technologies to “enhance, automate, or assist in the design process.” AI-driven design taps into AI algorithms, including machine learning and natural language processing, to … “solve complex design challenges.”

Given that broad definition, where does AI-powered design fit into the mix with generative design and topology optimization? Are the design concepts all interconnected? Further, how can AI-powered tools enhance the workflow? Several years back, DE’s Senior Editor Kenneth DE did a comparison of topology optimization to generative design. For this issue, DE consulted with several vendor experts to get a read on AI’s role and how it can impact the future of design.

Is AI-powered design just another name for generative design and topology optimization? If not, how would your company define it?

Fatma Kocer, Vice President Engineering Data Science, Altair (Kocer, Altair): AI-powered design is not just another name for generative design and topology optimization—they are very different approaches to design. They all use a combination of mathematical procedures but the main difference is that in the User-Defined column that includes generative design, design parameters are defined upfront. In AI-Powered Design, which is sometimes referred to as GenAI for Design, design space is learned from the data.

Ilya Tolchinsky, Ansys, Product Manager for AI/ML (Tolchinsky, Ansys): AI is a loaded term—different people are using it in different scopes. What we’re talking about today is machine learning, where we have a number of different algorithms. This is a branch of mathematics that has recently come up with novel technologies.

Efthymios Papoutsis, Technical Product Manager, Siemens Digital Industries Software (Papoutsis,Siemens): The term “AI-powered” is not on the same level as topology optimization and generative design. To explain with an example, topology optimization is a specific structural optimization method that optimizes the topology of a given boundary. There are different mathematical approaches on how to do this: level-set, density-based, etc.

For example, the density-based topology optimization method is an iterative process. In each iteration there is a series of operations that have to be done to propose a density field for the next iteration, continuing forth until convergence. These operations include a sensitivity analysis, an optimization step, etc.

My understanding of AI-powered is that AI methods such as neural networks are used in some or all of those operations. For example, in the case of sensitivity analysis, instead of calculating it numerically, a neural network can be trained to predict it faster.

Jon Hirschtick, Chief Evangelist, PTC (Hirschtick, PTC): AI-powered design is more than just another name for generative design and topology optimization. Generative design and topology optimization are just two of many possible uses of AI to power design.

What was it that generative design could do that topology optimization couldn’t have done? And what is it that AI-powered design can do that generative design couldn’t? 

Kocer, Altair: Generative design is the merger of topology optimization and design exploration. Design space exploration searches for different manufacturing requirements [and] engineering performance requirements so that the user knows what the trade-offs are, and hence makes a more informed decision.

In user-defined methods, [the] user identifies parameters based on their experience. In AI-powered design, we are fully learning from the data. This may add features that users miss or could not implement.

Papoutsis, Siemens: Generative design is an umbrella term and a process one level higher than topology optimization. Generative design is all those methods that can be used to generate shapes, from hanging chain and soap film models, all the way to structural optimization, metamaterials, and AI.

Hirschtick, PTC: Generative design can generate new shapes from merely specifying the goals and constraints of a design problem. Topology optimization starts with a designed part and then improves its shape. AI-powered design is, in my opinion, an umbrella term for a whole possible range of applications of AI, particularly modern AI based on large language models and diffusion-based imaging models.

Tolchinsky, Ansys: Generative design is a broader term that includes AI techniques for creating new geometries, which can be used in the context of optimization. Topology optimization is a particular type of technique for creating new geometries. It is a subset of generative design. Topology optimization algorithms can be AI powered.

In the search for a way to generate new shapes, we can use topology optimization, and as new AI techniques come up, we are using them now as another manifestation of a solution to the generative design problem.

Did anything serve as a catalyst for topology optimization to evolve into generative design, and generative design to evolve into AI-powered design? Did more powerful processors, GPU acceleration, or natural language processing help launch these transformations?

Kocer, Altair: Generative design is topology optimization run many times and hence HPC [high-performance computing] is definitely a catalyst. For AI-powered design, GPU and developments in geometric deep learning are catalysts. NLP is not a part of these processes.

Papoutsis, Siemens: This question is not very meaningful because topology optimization did not evolve to become GD, and GD did not evolve to become AI-powered design. The advancements in computing power have paved the way for AI becoming popular and powerful, and naturally its usefulness has been and is being researched in topology optimization and other generative design methods.

Hirschtick, PTC: Generative design and topology optimization grew up more or less together as applications. Other AI-powered design applications are catalyzed by the recent hypergrowth in the capabilities of large language models and AI-based imaging technology, which in turn rely on GPU hardware. Another key factor: cloud-native applications, which have obvious benefits both in the large data and large compute aspects of AI-powered design.

Any current/future applications of AI-powered design you’d like to share?

Tolchinsky, Ansys: At Ansys, we’re working on something where a user can work in their traditional way. The designer comes up with the geometry, evaluates its performance and, using their experience, modifies the design for the next variant and continues this process.

We’re working on an AI that’s observing them doing this and looking at [what] they’ve tried. Then [we’re] suggesting a new shape for their design based on what they’ve done so far. This Copilot assistant allows the user to work in their natural way but is using AI, optimization and generative design to really enhance the workflow.

(As for) Ansys tools with AI-powered capabilities, what we have is still in development.

It’s coming this year—beta should [be released] in mid-year.

Hirschtick, PTC: Potential applications we are working on at PTC include AI-powered design advisors, or copilots, and AI-powered design rendering. AI copilots would give expert-level advice to designers on how to approach a problem in computer-aided design. AI-powered design rendering applications would provide a rendered image of a design, including filling in portions of the design not yet completed, based on merely a user’s CAD model and/or textual instructions.

Daniel Graham, Senior Director of Product Management, Autodesk (Graham, Autodesk): What we found is, if you build a CAD tool, the only automation you can drive is within the scope of that CAD tool. If we have upfront design, mechanical design, validation, manufacturing, the cloud, data management for all of it, think about the kinds of automation that we can do—all the way through the process.

But to do that … in architecting Autodesk Fusion, we had to converge what used to be separate applications, and the number one point is cloud convergence. We have to unify them into an end-to-end product development package.

Second, we had to have cloud data. We did a lot of work so that we could drive automation through this. So now we have a common compute stack.

In terms of incorporating AI into this, [we did that] within the past five years. We started the unification of the application. We built it on a cloud—a native data stack. We’ve looked for opportunities to add really practical AI examples into our customer workflows. Let’s just take the task that they already have and … make that automated.

Jeremy Stadtmueller, Director of Product Management, Autodesk (Stadtmueller, Autodesk): You have to understand how CAD works, to set up the problem and then solve for it. There are thousands, if not millions, of users out there that spend time learning CAD, and now this will optimize for that. This says, “You’re good at this. I can make you great at it.” But there’s this fundamental gap between a non-user and a user. That’s where AI can really come in.

Whenever you create a solid model, you have to sketch something in 2D. That sketch needs dimensions, constraints. Even if you were just an artist in drawing, you might inherently understand or have been trained [about those dimensions and constraints]. But from an engineering standpoint, those things staying related and being updated, it requires training an update. And then there’s software. So the software has to be stable and wants to know the specific definition of every little detail to understand everything. There’s a lot of detail that this sketch needs, that the computer needs but the user doesn’t care about. So we’re using AI tools to look at millions of sketches. Then, because they’re large language models—foundation generative models that we build—and say, “This is how sketches are constrained” and we’ll automatically add a tangency, a dimension, or whatever else you need to define the sketch. And a user can say, ‘Well, I really just need it to be a length and a width. Everything else is immaterial to me.’ The system will just figure all that out for you.

This released in January of 2024 and is called Sketch AutoConstrain. Our users don’t care if it’s AI-powered or not. They just want it automated. They just want it to work.

Any additional perspective you’d like to provide?

Tolchinsky, Ansys: In most of what we do, we are talking about training AI models from scratch. The AI model knows nothing. The user shows them some of their own data and then they own everything it produces. It’s encapsulated and compartmentalized—there is isolation between customers.

As a simulation vendor if you look at what we’re trying to do with our products, [we want to] make the simulation faster, more accurate, easier to set up. AI has helped us to take a giant leap in this direction. But this just gets to where we were already going faster. What engineers really want is for AI to help us find designs that are beyond the limits of human imagination.

Stadtmueller, Autodesk: Our perspective is that AI is not a value for AI’s sake. It is: can we make their lives better? Can we make their workflows [run] faster? That’s why it can’t be so complex that they’re [saying], ‘I don’t want to have to learn an entirely new thing right away.’ That’s why we start with these small automations that make their lives easier and to create proof and trust that this is a system that works.

Graham, Autodesk: We are a good way into this [AI] journey. I think one of the things that we’re going to see is it will be the norm in other tools that users use. They’re going to be exposed to AI in other parts of their workflow, and I think that there’s going to be a hunger and an opportunity for them to see the same kind of value in their product development process.

More Autodesk Coverage