Generative CAD & AI Surrogate Models: Closing the Loop Between Design and Simulation

For most of engineering history, designing a physical product has meant alternating between two slow, separate worlds: building geometry in CAD, and then waiting — sometimes hours, sometimes days — for a simulation to tell you whether that geometry actually works. Two threads of AI research are now quietly collapsing that gap: generative models that produce CAD geometry directly, and surrogate models that approximate simulation results almost instantly. Individually, each is useful. Together, they turn engineering design into something closer to a real-time, interactive search process.

The Problem with Traditional CAD

Conventional CAD is manual by design. An engineer sketches, extrudes, fillets, and booleans their way to a shape, one command at a time. Even "generative design" tools that have existed for years — like topology optimization, where an algorithm removes material from a design space based on load conditions — tend to output raw meshes or point clouds. These looked like optimized parts, but they weren't CAD models in the sense that matters to an engineer: parametric, editable, and expressed as a history of construction operations that a downstream tool or colleague could modify.

That distinction — geometry versus an editable construction history — is the central problem a newer generation of machine learning research has tried to solve.

Teaching Machines to "Draw" in CAD's Own Language

The first serious attempt to get a neural network to generate actual CAD command sequences, rather than raw shapes, was DeepCAD (Wu et al., ICCV 2021). The insight was to treat CAD construction the way language models treat text: a CAD model is really just a sequence of discrete operations — sketch this curve, extrude it this far, cut this pocket — so a transformer architecture built for sequential data could, in principle, learn to generate new sequences of its own. To train it, the researchers built a dataset of over 178,000 real CAD models and their construction sequences, scraped from the Onshape platform, and released it publicly for others to build on.

The following year, a team at Autodesk Research pushed the idea further with SkexGen (Xu et al., ICML 2022), which focused specifically on "sketch-and-extrude" style CAD — the dominant modeling pattern in mechanical design. SkexGen's contribution was disentanglement: instead of encoding a whole design into one blob of latent information, it split the representation into separate codebooks for topology, geometry, and extrusion. That separation meant a designer (or an algorithm) could take the topology of one design, the geometry of another, and combine them into something new — a kind of controllable remixing of CAD models that plain autoencoding doesn't allow.

Both papers point at the same underlying goal: geometry that isn't just visually plausible, but usable — something you could open in a real CAD package, edit a parameter on, and regenerate.

The Other Half: Simulating Without Simulating

Generating geometry faster is only half the problem. The other bottleneck is finding out whether that geometry is any good — structurally, aerodynamically, thermally — which traditionally means running finite element analysis (FEA) or computational fluid dynamics (CFD), both of which can take anywhere from minutes to days depending on complexity.

Surrogate models sidestep this by training a network to approximate the input-output relationship of a simulation directly, using a dataset of prior simulation runs. Three architectural approaches have defined this space:

MeshGraphNets, from DeepMind (Pfaff et al., ICLR 2021, awarded an outstanding paper prize), applies graph neural networks directly to a simulation mesh, passing messages between connected mesh nodes the way a solver would propagate stress or pressure between neighboring elements. Crucially, the model can also adapt the mesh's resolution as it runs, concentrating detail where the physics demands it. The results were validated across aerodynamics, structural mechanics, and cloth simulation, and the model ran one to two orders of magnitude faster than the traditional solver it was trained to imitate.

Fourier Neural Operators (FNO), developed by Zongyi Li and collaborators, take a different approach: rather than learning on a mesh, they learn a resolution-invariant operator — a mapping between entire function spaces — using Fourier transforms as the core computational layer. The original formulation was fast and accurate on problems like turbulent fluid flow, but was restricted to simple rectangular domains with regular grids, since it relied on the standard fast Fourier transform. A follow-up variant, Geo-FNO, solved this by learning a coordinate deformation that maps irregular, real-world geometries onto a regular computational grid before applying the FFT — meaning the model can accept point clouds, meshes, or raw design parameters as input, not just clean rectangular domains.

DeepONet, from Lu Lu, George Karniadakis, and collaborators, took yet another route, grounding its architecture in the universal approximation theorem for operators. It uses a "branch net" to encode the input function and a "trunk net" to encode the query location, combining them to predict a solution — a formulation that has since been extended into physics-informed and multi-fidelity variants for problems ranging from bubble dynamics to nanoscale heat transport.

What unites all three is the same trade: a large upfront investment in generating training data from real simulations, in exchange for near-instant predictions afterward — often fast enough to sit inside an interactive design tool rather than a batch queue.

Where the Two Threads Meet

The genuinely interesting part isn't either technology alone — it's what happens when you wire them together into a loop:

1. A generative model proposes a candidate geometry.
2. A surrogate model estimates its performance in milliseconds instead of minutes or hours.
3. That fast feedback steers the next round of generation — whether through gradient-based optimization, evolutionary search, or reinforcement learning.

This loop is what separates "AI-assisted design" from the older topology-optimization tools that already existed. Traditional optimization loops were bottlenecked by the simulation step; each iteration meant another full FEA or CFD run. Replacing that step with a surrogate model — one that's already been trained to approximate the same physics — means the loop can run thousands of times in the time it used to take to run once. The generative side supplies genuinely novel candidate shapes instead of just perturbing an existing one, and the surrogate side tells you almost instantly whether a given candidate is worth pursuing further.

Industry Is Already Building Toward This

This isn't purely academic. nTop (formerly nTopology), a computational design software company, has been assembling exactly this kind of pipeline through acquisition. In February 2025, nTop acquired cloudfluid, a German company whose GPU-native CFD solver predicts fluid flow without requiring the complex conformal meshing that traditional CFD tools demand. Paired with nTop's existing implicit geometry kernel — which represents shapes mathematically rather than as explicit meshes — the combination lets engineers iterate on fluid-dynamics-informed designs in near real time rather than waiting on conventional solvers. nTop has also been integrating with NVIDIA's accelerated computing platforms to push more of this computation onto GPUs.

It's a concrete example of the generative-CAD-plus-surrogate-model loop moving from research papers into commercial engineering software, aimed at industries — aerospace, medical devices, automotive — where iteration speed on physically valid geometry is a genuine competitive advantage.

What's Still Unsolved

None of this is a finished story. A few open problems are worth naming:

• Editability still isn't guaranteed. Even sequence-based generators like DeepCAD and SkexGen can produce invalid or self-intersecting geometry, and translating a generated sequence back into a fully editable model in a mainstream CAD package remains nontrivial.
• Surrogate models are only as good as their training distribution. A network trained on one class of geometries or boundary conditions can fail silently — producing a confident but wrong answer — when asked to extrapolate to genuinely novel designs, which is exactly the situation a generative model is likely to create.
• Trust and validation. Engineering sign-off, especially in safety-critical industries, still generally requires a final pass through a conventional, physics-based solver. Surrogate models are best understood today as an accelerant for exploration, not yet a replacement for certification-grade analysis.

Even with those caveats, the direction is clear. The distance between "I have an idea for a shape" and "I know whether that shape actually works" has been shrinking for several years, and the combination of sequence-based generative CAD with graph- and operator-based surrogate modeling is the most concrete technical path toward closing it further.