Neural CAD: what it means and who's building it

Neural CAD means neural networks that generate CAD modeling operations, sketches, extrusions, fillets, the actual construction sequence, rather than predicting final geometry. Autodesk coined the term for their own research effort, but the idea is broader than one company. The goal is AI that thinks in feature trees, not triangulated surfaces. It's the difference between an AI that draws a picture of a bracket and an AI that tells you how to build one step by step in Fusion 360. The research is genuine. I've read the papers. The production readiness, as of April 2026, is not there.

I first heard the phrase "Neural CAD" during Autodesk University 2025, watching the keynote stream on my laptop with a cup of coffee that went cold before the demo was half over. Mike Haley from Autodesk Research used the term while showing a prototype that generated editable geometry inside Fusion 360's canvas from a text prompt. The audience applauded. I wrote "when?" in my notes and underlined it twice. Six months later, the answer is still "not yet," but the underlying research has continued moving, and it's worth understanding what's actually happening beneath the marketing.

What "neural CAD" means technically#

Traditional text-to-3D AI, the kind that powers tools like DreamFusion or Point-E, generates geometry as a final output. You give it text, it gives you a shape. The shape might be a point cloud, a mesh, a NeRF, or a voxel grid. What it isn't is a construction history. There's no sequence of operations you can replay, edit, or modify. The output is dead geometry. You can look at it, but you can't really work with it.

Neural CAD flips this. Instead of predicting what the final geometry looks like, the network predicts the sequence of CAD operations that produce the geometry. Sketch a rectangle on the XY plane. Constrain it to 80mm by 50mm. Extrude it 10mm in the positive Z direction. Sketch a circle on the top face. Constrain it to 5mm diameter. Cut-extrude through all. That kind of sequence.

This is fundamentally harder. Predicting a surface is a regression problem: approximate the right coordinates. Predicting a construction sequence is a structured prediction problem: get the right operations, in the right order, with the right parameters, and have the whole thing compile into valid geometry when executed by a CAD kernel. It's like the difference between predicting what a finished house looks like versus predicting the construction plan that builds it. The plan is more useful, but it's also more constrained and more brittle when things go wrong.

The how text-to-CAD works post covers the broader technical landscape, but the neural CAD approach represents the most ambitious branch of that tree. It's aiming not just for geometry, but for engineering-grade output.

The key research#

Three lines of research define the neural CAD space as of 2026. They're related but distinct.

DeepCAD, published in 2021, introduced the dataset and the autoencoder architecture that most subsequent work builds on. The DeepCAD dataset contains approximately 178,000 parametric CAD models from ABC (a large CAD dataset derived from Onshape public models), represented as sequences of sketch-and-extrude operations. The DeepCAD model learned to encode these sequences into a latent space and decode them back into valid operation sequences. It proved that neural networks could learn the language of CAD construction, not just the appearance of CAD geometry.

The dataset itself is a major contribution. Before DeepCAD, there was no large-scale collection of CAD models stored as operation sequences with enough variety to train a neural network. Image generation had ImageNet and LAION. Language models had the internet. CAD generation had almost nothing. DeepCAD created the foundation, limited as it is (the models are geometrically simple, mostly prismatic parts with basic features).

The Text2CAD paper, published as a NeurIPS 2024 spotlight, built on DeepCAD by adding text-conditioning. The team annotated the DeepCAD dataset with approximately 660,000 text descriptions at multiple skill levels (beginner, intermediate, expert) and trained a transformer that takes a text prompt and generates a CAD operation sequence. This was the first end-to-end pipeline from natural language to parametric CAD. The model architecture uses a BERT encoder for text and an autoregressive decoder for CAD tokens, predicting each operation conditioned on the text encoding and the operations generated so far.

Autodesk's Neural CAD research, presented at AU 2025, takes a proprietary approach that builds on similar ideas but with Autodesk's internal data and engineering. The details are thinner because Autodesk hasn't published the architecture with the same openness as academic work. What they've shown publicly is a foundation model trained on CAD geometry that can generate native, editable B-Rep geometry inside Fusion 360. The demo at AU showed text-to-geometry generation producing objects with selectable faces and edges in the Fusion canvas. The Fusion 360 Neural CAD post covers what Autodesk has said publicly about this effort.

Why generating operations is harder than generating meshes#

The difficulty gap between generating a mesh and generating a CAD construction sequence is enormous, and understanding why explains a lot about the current state of neural CAD.

A mesh is forgiving. If a triangle is slightly wrong, the overall shape might still look fine. Meshes degrade gracefully. You can have a mesh that's ugly up close but perfectly usable at a reasonable zoom level. And there's no requirement for internal consistency beyond the faces connecting at edges. This is why text-to-3D tools that generate meshes have improved so rapidly. The problem is inherently tolerant of small errors.

A CAD operation sequence is not forgiving. If a sketch constraint is wrong, the sketch might be invalid. If an extrusion references a face that doesn't exist because a previous operation failed, the whole sequence breaks. If the parameters are slightly off, you don't get a slightly wrong part. You might get an error, or a completely different part, or a kernel crash. CAD sequences are like programs: they either compile or they don't, and a single wrong token can break everything.

This brittleness is why neural CAD models produce a meaningful percentage of invalid outputs. The Text2CAD paper reports invalidity ratios for different generation conditions, and while the valid outputs are impressive, the invalid ones highlight the core challenge. Generating a sequence that looks statistically correct is not the same as generating a sequence that executes correctly in a geometric kernel. Every operation has preconditions, and the network has to satisfy all of them in order.

There's also the combinatorial problem. The number of possible mesh surfaces for a given shape is large but continuous. You can interpolate between nearby meshes smoothly. The number of valid CAD operation sequences for a given shape is discrete and combinatorial. A bracket can be built a hundred different ways: sketch the L-profile and extrude, or sketch a rectangle and extrude twice, or use a shell operation on a box. Each construction yields the same visual result but a different feature tree. The network has to pick not just the right geometry but a valid construction strategy, and there are many valid strategies with no clear way to smoothly interpolate between them.

The B-Rep advantage#

The reason neural CAD matters, the reason researchers are going through this pain instead of just generating meshes, is B-Rep.

B-Rep (Boundary Representation) is how professional CAD tools represent solid geometry internally. Faces, edges, vertices, and the topological relationships between them. A B-Rep model has real faces you can select, real edges you can fillet, real surfaces you can measure. It's the native language of manufacturing. STEP files are B-Rep. SolidWorks files are B-Rep. Every CNC machine and every mold shop works from B-Rep geometry.

Meshes are approximations. A mesh representation of a cylinder isn't a true cylinder. It's a collection of flat triangles arranged to look like a cylinder. You can't grab a face of a mesh cylinder and extrude it cleanly. You can't add a precise fillet to a mesh edge. You can't measure a mesh dimension and get an exact number. For engineering work, mesh output from an AI is fundamentally limited.

Text-to-CAD tools like Zoo.dev already generate B-Rep output, but they do it through their own geometric kernel (KittyCAD) rather than through neural operation-sequence prediction. The output is B-Rep, which is great for manufacturing, but it doesn't have an editable feature tree. You get a solid body. You can measure it and machine from it. You can't easily modify it by changing a dimension in a timeline.

Neural CAD is after the full prize: B-Rep geometry with an editable construction history. A model that comes out of a neural CAD system would be indistinguishable from one a human modeled in Fusion 360. You could roll back the timeline, change a sketch dimension, and watch the rest of the model update. That's the goal, and nobody has delivered it at production quality yet.

Current state of production readiness#

The honest assessment of neural CAD in April 2026:

The research works. The Text2CAD paper demonstrated that text-to-CAD-operations is possible. DeepCAD demonstrated that neural networks can learn to generate valid construction sequences. Autodesk has demonstrated internally that this can work inside a real CAD environment.

The output quality is not production-grade. The models in the DeepCAD training set are simple. Boxes, cylinders, plates, basic mechanical shapes. Neural CAD models generate geometry within that vocabulary. Ask for a complex injection-molded housing with snap-fits, ribs, and draft angles, and you're far outside the training distribution. The dimensional accuracy is approximate. The feature vocabulary is limited (sketch-and-extrude dominates; fillets, chamfers, patterns, and sweeps are mostly absent from the training data).

No commercial tool offers neural CAD output to end users. Autodesk has announced it. They've demoed it. They haven't shipped it. The Fusion 360 AI features inventory makes this clear: Neural CAD is on the roadmap, not in the product. The Text2CAD research code is available on GitHub but is non-commercial and research-grade.

The training data problem is unsolved. Image generation models train on billions of images scraped from the internet. The largest CAD operation sequence dataset (DeepCAD) has 178,000 models. That's a five-order-of-magnitude gap. Real engineering CAD data is locked inside companies, protected by IP concerns, and stored in proprietary formats that are hard to extract operation sequences from. Until this data gap narrows, neural CAD models will be limited to simple geometry.

What it means for the future#

Neural CAD is the path to AI-generated models that engineers can actually edit. Not just view, not just manufacture from, but open in Fusion 360 or SolidWorks and modify as if they'd built it themselves. That's the parametric AI design goal that the whole field is chasing, and neural CAD is the most plausible technical approach to getting there.

The timeline is unclear. Autodesk has the research team, the data (Fusion 360 has millions of models created by users), and the motivation. But the gap between "works in a demo" and "ships in a product" has historically been measured in years for this kind of technology. Fusion 360's generative design took years to go from research to stable product. Neural CAD is a harder problem.

The text-to-CAD guide covers what's usable today, and none of the currently shipping tools use neural operation-sequence generation as their primary approach. Zoo.dev generates B-Rep through a different method. CADAgent uses LLMs to issue Fusion 360 API commands, which is closer to neural CAD in spirit but architecturally different. The pure neural-CAD approach, where a single network predicts a complete construction sequence from text, is still in the lab.

My expectation is that when neural CAD does ship commercially, it'll start narrow. Simple parts, limited feature vocabulary, modest dimensional accuracy. The way autocomplete started with word suggestions and gradually became paragraph-level generation. The first production neural CAD tool won't replace a senior CAD engineer. It'll replace the first five minutes of modeling a simple part, and it'll do it with a feature tree instead of a dead body. That alone would be worth the decade of research it took to get there.

For now, neural CAD is the most interesting thing happening in AI-generated geometry research, and the least useful thing for someone who needs a bracket by Thursday. Both of those statements are true, and there's no contradiction. The research is building the foundation. The production tools will come later. I'm watching it closely and modeling my brackets the old way in the meantime.