Here is what actually issued. On June 16, 2026, the U.S. Patent and Trademark Office granted Autodesk, Inc. U.S. Patent No. 12,657,468, “Techniques for training machine learning models to automate tasks associated with 3D CAD objects.” It is a grant — kind code B2 — not a pending application, and the distinction matters: the rights are live as of the issue date. What is covered is a training method, and unlike a lot of “AI for design” filings, this one is classified at the core of machine learning rather than at the edges.
The independent claim is procedural, and it reads like the training loop it describes. A training application computes a preliminary result via a machine learning model based on a representation of a 3D CAD object — a representation that includes a graph and multiple 2D UV-grids. Based on that preliminary result, the application performs operations to determine that the model has not yet been trained to perform a first task. It then updates at least one parameter of a graph neural network included within the model, based on the preliminary result, to generate a modified model. Finally, it performs operations to determine that the modified model has now been trained to perform that first task. That is a supervised training cycle stated in claim terms: infer, measure the gap, adjust the GNN's weights, confirm convergence.
"The training application updates at least one parameter of a graph neural network included in the machine learning model based on the preliminary result to generate a modified machine learning model."— U.S. Patent No. 12657468, source
The named inventors are Pradeep Kumar Jayaraman, Thomas Ryan Davies, Joseph George Lambourne, Nigel Jed Wesley Morris, Aditya Sanghi, and Hooman Shayani. The CPC classification is unusually clean for a CAD patent: the lead symbol is G06N 3/084, which is backpropagation in neural networks — the training mechanism itself. It carries a full stack of G06N 3 companions: G06N 3/045 (networks built from sub-networks), G06N 3/0464 (convolutional architectures), G06N 3/088 (unsupervised learning), and G06N 3/09 (supervised learning). The CAD context shows up in G06F 30/10 and G06F 30/27 (computer-aided design, and CAD using machine learning), and the graph structure is reflected in G06F 16/9024 (graph databases). This is a machine learning patent that happens to be about CAD, not a CAD patent that name-drops AI.
Why a graph neural network for B-reps
The representation choice is the whole story. A boundary representation, or B-rep, is how solid 3D models are actually stored in CAD systems: a model is described by its faces, edges, and vertices and the topological relationships connecting them. That structure is a graph by nature — faces are nodes, shared edges are connections — which makes a graph neural network a natural fit in a way a plain convolutional network applied to a rendered image is not. The patent's representation pairs that topological graph with multiple 2D UV-grids, the parameterized surface coordinates that capture each face's geometry. So the GNN learns over the topology while the UV-grids carry the local surface shape. Together they give a model that can reason about a CAD object the way the CAD kernel stores it, rather than the way a camera sees it.
That is the deflationary but real value here. There has been a wave of attempts to apply learning to CAD by converting solids into point clouds, meshes, or voxels — all of which throw away the exact topological and parametric information the B-rep already contains. Training a graph neural network directly on the B-rep graph and its UV-grids keeps that information intact. The claimed loop, where the training application checks whether the model has learned a task and updates GNN parameters until it has, is a generic-sounding supervisory wrapper, but it is recited specifically around the graph-plus-UV-grid representation, and that is what bounds the claim.
What the grant covers, and how narrowly to read it
Read claim 1, then talk scope. What issued is a training methodology tied to a particular CAD-object representation and a graph neural network — not a monopoly on machine learning for design, nor on graph neural networks generally. The novelty Autodesk is asserting lives in the combination: the B-rep-derived graph and 2D UV-grids as the model's input, a GNN inside the model, and a training application that iterates on the GNN's parameters until a target task is learned. Strip out the specific representation and you are left with ordinary backpropagation, which is why the claim leans on that representation to carry its weight.
For Autodesk, the fit is strategic and on-brand. The company's entire franchise is CAD and design software, and a grant that covers how to train models to automate tasks on the exact B-rep data its products already produce is a defensive and offensive asset for the design-automation features it is building. The roster of six inventors, several with research backgrounds in geometric deep learning, signals this came out of a deliberate program rather than a one-off. Granted, not just filed: as of June 16, this is an enforceable piece of Autodesk's growing machine-learning-for-CAD estate, classified right where it belongs — under G06N 3.