CADChain Blog

Machine Learning in CAD in 2026: How AI Is Changing Engineering Design, From Self-Constraining Sketches and Automated Drawings to Text-to-CAD and Physics AI | CADChain | CAD DRM & IP

Guides AI/ML
For most of its history, CAD software did exactly what the engineer told it to do: draw this line, extrude this profile, add this fillet. Machine learning in CAD changes that relationship. The software now learns patterns from past sketches, parts, drawings and simulations, and starts to suggest the next step: the constraints a sketch is missing, the views a drawing needs, the command you will probably use next, or how a new shape is likely to perform. This guide explains what machine learning in CAD really is, how it works on CAD data, where it already shows up in tools such as Autodesk Fusion, Siemens NX, SOLIDWORKS, Onshape and BricsCAD, what physics AI adds to simulation, where the limits are, and the intellectual property questions every engineering team should ask before AI touches its design files.
TL;DR: Machine learning in CAD means models trained on design data (sketches, B-rep models, meshes, drawings, simulation results) that predict something useful inside the design workflow. In 2026 the most mature uses are narrow and practical: auto-constraining sketches, generating 2D drawings, predicting the next command, answering how-to questions, cleaning up drawings and predicting simulation results in seconds. Text-to-CAD and generative 3D models exist but are early. ML suggestions still need an engineer to check them, the models are only as good as the data they learned from, and every team should know where its CAD files go, whether they train someone else's model and how it will prove which designs are its own.

What is machine learning in CAD?

Machine learning is a branch of artificial intelligence in which a model learns patterns from examples instead of following rules a programmer wrote by hand. In computer-aided design, the examples are design data: thousands of sketches with their constraints, 3D parts with their modelling history, drawings with their views and dimensions, assemblies with their joints, or simulation runs with their results. Once trained, the model can make a prediction about a design it has never seen, such as which dimensions a new sketch needs or how air will flow around a new body shape.
Three terms often get mixed up, so it is worth separating them early:
  • Rule-based automation (macros, design tables, configurators, classic auto-dimensioning) follows logic someone wrote. It is predictable and does not learn.
  • Generative design and topology optimisation compute shapes from the loads, materials, keep-out zones and manufacturing methods you set. They are often marketed as AI, but they work from your problem definition, not from a library of past designs.
  • Machine learning, including today's generative AI and large language models, learns from data and then predicts. That is what makes it powerful, and it is also why its training data, its errors and its ownership questions matter.
CADChain has written about narrower uses before, such as machine learning for CAD file access pattern analysis and AI-powered CAD file anomaly detection, which use ML to protect design files. This article is about ML inside the design work itself.

How does machine learning work on CAD data?

CAD data is harder for machine learning than text or photos. A part is not a flat list of words or pixels: it is a boundary representation (B-rep) of faces, edges and vertices with exact geometry, plus a feature history, constraints and design intent. Researchers turn that into something a model can read in a few ways:
  • Sequences. A sketch or a modelling history becomes a sequence of tokens (line, arc, coincident, distance), the way a sentence is a sequence of words. This is how many language-model techniques reach CAD.
  • Graphs. The faces of a B-rep model become nodes and their shared edges become links, so a graph neural network can classify faces, find features or compare parts.
  • Meshes, point clouds and voxels. Simpler 3D formats that many vision and generative models already understand, at the cost of losing exact, editable geometry.
  • Images of drawings. 2D drawings and scans, for title-block reading, symbol recognition or conversion to vector data.
Training data is the hard part. Two public research datasets show what it looks like. The ABC dataset collects one million CAD models provided by Onshape, and its page states that the copyright of the models is owned by their creators. Autodesk's Fusion 360 Gallery dataset, built from designs shared in the Autodesk Online Gallery, includes 8,625 sketch-and-extrude construction sequences, 35,680 parts labelled by the modelling operation that created each face, and 8,251 multi-part assemblies with joint information. Commercial vendors train on their own data under their own terms, which is why the terms matter (more on that below).
Autodesk's AutoConstrain is a good example of the whole loop. According to Autodesk Research's research post on AutoConstrain, it uses a Transformer model trained on a large set of constrained sketches: some constraints and dimensions were removed, and the model learned to predict them back. Inside Fusion, the same model proposes constraints and dimensions for a sketch the user is drawing.

Where is machine learning used in CAD software today?

The useful applications in 2026 are less dramatic than "AI designs the product". They save minutes on tasks engineers repeat every day. The table lists real examples, each checked on the vendor's or research team's own page.
Tool or projectType of machine learning useWhat it does for the engineerStatus
Autodesk Fusion AutoConstrainSketch constraint prediction (Transformer)Proposes constraints and dimensions to define a sketchReleased in the January 2025 Fusion update for commercial subscribers
Autodesk Fusion automated drawingsDrawing generationCreates 2D views and dimensions from a 3D model and leaves out fastenersAvailable in Fusion
Siemens NX adaptive UICommand predictionLearns from user actions and shows the commands you will probably need nextPart of NX (Siemens Designcenter)
SOLIDWORKS 2026 AURA and LEOAI virtual companionsAURA gives design guidance and captures knowledge from 3DSwym contentIn SOLIDWORKS 2026 through the SOLIDWORKS Labs (Beta) tab
Onshape AI AdvisorAssistant trained on product knowledgeAnswers how-to questions from Onshape documentation and best practicesInside Onshape and the Onshape Learning Center
BricsCAD Blockify and BIMifyPattern recognition in drawings and modelsFinds similar geometry to turn into blocks; classifies BIM elementsIn BricsCAD (Bricsys, now part of Octave)
Ansys SimAIPhysics AI surrogate modelLearns from past simulation results to predict how new designs performCommercial product (Ansys, part of Synopsys)
NVIDIA PhysicsNeMoFramework for physics AILets teams build, train and run their own physics AI modelsOpen source
Zoo Design Studio (Zookeeper, formerly Text-to-CAD)Text-to-CAD agentCreates editable 3D models from a text promptCommercial, early stage
Autodesk Project BerniniGenerative 3D shapesGenerates 3D shapes from text, images, point clouds or voxelsResearch project, trained on ten million 3D shapes

Six ways machine learning is changing engineering design

1. Sketches that constrain themselves

Under-constrained sketches are one of the most common reasons models break when a dimension changes. ML constraint prediction, such as AutoConstrain in Fusion, proposes the missing relations and dimensions so the engineer can accept a fully defined sketch in one step instead of adding them one by one. The engineer still chooses which dimensions carry the design intent.

2. Drawings generated from the 3D model

Creating 2D manufacturing drawings is slow and repetitive. Fusion's automated drawings feature analyses the 3D model, generates the views and dimensions, and leaves out items such as standard fasteners. The result is a first draft that a drafter checks and finishes, not a signed-off drawing.

3. Interfaces and assistants that learn how you work

Siemens describes the NX adaptive user interface as an assistant that uses AI and machine learning to track user actions, predict the commands you will likely use next and show them in a compact panel that adapts to the task, such as sheet metal work. Chat-style assistants go further: Onshape AI Advisor answers questions from Onshape's own documentation, and SOLIDWORKS 2026 added the AURA and LEO virtual companions. Agents that act inside CAD are the next step, which we covered in where a CAD AI agent fits between design brief and 3D prototype. Our sister sites follow the same trend for what long-running AI companions such as LEO mean for SOLIDWORKS data and for Autodesk Inventor AI after AU 2026.

4. Shapes from text, images and scans

Generative models can now produce 3D geometry from a prompt or an image. Zoo's Zookeeper, the successor to its Text-to-CAD tool, creates editable 3D models from a prompt inside Zoo Design Studio. Autodesk Research's Project Bernini, an experimental model trained on ten million diverse 3D shapes, generates shapes from text, images, point clouds or voxels. Research is also turning meshes back into CAD, as in our article on AI that rebuilds STL meshes as editable CAD. For mechanical parts these tools are early: expect concept geometry to refine, not production-ready parts. They also raise new generative AI and CAD IP protection challenges.

5. Physics AI: simulation results in seconds

Classic simulation solves physics equations for every new design, which can take hours. Physics AI models learn from earlier simulation results and predict the result for a new design almost instantly. According to Ansys's SimAI product page, SimAI learns from existing simulation data, typically needs 30 to 100 simulation results to produce a model with sufficient accuracy, predicts a continuous 3D physical field and gives each result a confidence level that shows whether the design is too far from the training data. NVIDIA PhysicsNeMo is an open-source framework for teams that want to build and train such models themselves. Physics AI is best for exploring many variants quickly; final verification still belongs to a full simulation or a physical test.

6. Cleaner drawings, reuse and design checks

Pattern recognition helps with the unglamorous work. BricsCAD's Blockify detects similar geometry in a drawing and groups it into blocks, and BIMify classifies building elements. The same family of techniques powers similar-part search across a vault, automatic feature recognition and checks that flag a design that looks unlike anything that passed review before. On the security side, the same ideas are used to spot unusual access to design files, and federated learning for collaborative CAD security lets partners train models without pooling their files.

Machine learning vs generative design vs generative AI in CAD

The terms overlap in marketing, but they answer different questions:
  • Generative design asks: given these loads, materials, keep-out zones and manufacturing methods, which shapes meet my goals? Autodesk describes generative design in Fusion as producing multiple manufacturing-ready solutions from manufacturing constraints and performance requirements, with a solid model you can keep editing.
  • Machine learning asks: given everything this model has learned from past data, what is the likely answer here? Constraint prediction, command prediction and physics AI all fit this pattern.
  • Generative AI is machine learning that creates new content, such as a 3D shape from a prompt, a drawing or a macro. It is the newest and least predictable of the three.
In practice, tools increasingly combine them: an ML model might predict performance for thousands of generative design candidates, or a language model might set up the generative design study for you.

What are the limits of machine learning in CAD?

  • Data quality. A model learns the habits in its training data. Half-constrained sketches, outdated standards or messy legacy models teach it the wrong lessons. Autodesk's own AutoConstrain post notes that not all training sketches were fully constrained, which the team had to correct for.
  • Precision. Engineering needs exact geometry and tolerances. Many generative 3D models produce meshes or approximate shapes that are hard to edit or manufacture.
  • Out-of-range designs. A model is most reliable close to what it has seen. That is why SimAI reports a confidence level and why a radically new design still needs full simulation.
  • Accountability. ML suggestions have no professional responsibility. A named engineer has to check and sign off every drawing, constraint scheme and predicted result.
  • Security. AI add-ins, agents and model files are new software in your toolchain. Our article on AI supply chain attacks on CAD and PLM toolchains shows why they need the same scrutiny as any plug-in, alongside known risks such as malicious DWG and DXF files.

Machine learning, CAD data and intellectual property

Machine learning runs on design data, and design data is often a company's most valuable intellectual property. Before you switch on an AI feature or upload a model to an AI tool, ask where the file goes, whether it can be used to train a shared model, who owns the generated output and how you will prove which parts of a design are yours. Many CAD files are trade secrets, and secrecy only protects you if you can show you took reasonable steps to keep them secret, as our trade secrets guide for deep tech startups explains.
Ownership of AI-assisted output is still being worked out by courts and patent offices; our article on AI-generated inventions and who owns them looks at what engineering teams should record. In Europe, the AI Act adds obligations for AI providers and deployers, discussed in EU AI Act enforcement and what it means for your CAD data. Artists face the same questions about AI training opt-out signals.

How to start using machine learning in your CAD workflow

  1. Pick one painful, repetitive task. Sketch constraints, drawing creation, part search or early simulation are good first candidates, because the benefit is easy to measure.
  2. Try the ML features already in your CAD tool. Fusion, NX, SOLIDWORKS, Onshape and BricsCAD all ship some ML today, with no data science project needed.
  3. Read the AI terms before you upload anything. Check where data is processed, whether it trains shared models, whether you can opt out and who owns the output.
  4. Clean the data you will learn from. For custom models or physics AI, curate fully defined, current, correctly labelled designs and simulation runs.
  5. Keep a human check in the loop. Name who reviews AI output, and keep that review in your normal release process.
  6. Protect the files before and after AI touches them. Use access control, encryption and rights management for sensitive models. Our CAD file security guide covers the basics, with more detail on CAD encryption and on information rights management for engineering files.
  7. Record what is yours. Register dated proof of your designs so you can show what existed before and after AI assistance.

Where CADChain fits: protecting the designs that train and use AI

Machine learning makes design data more valuable, because data now trains models as well as making products. That raises two questions CADChain was built to answer. Who can access your CAD files, including the copies that go to suppliers, partners and AI tools? And can you prove that a design was yours, and from when? CADChain combines file-level access control for shared designs with a blockchain record of ownership, so a team can keep control of its models and keep dated evidence of its work. Read more about blockchain ownership registration for CAD designs and the legal proof requirements for CAD file ownership.
Use AI on your designs without losing control of them. CADChain adds access control for shared CAD files and a dated record that the designs are yours. Discover CADChain

Conclusion

Machine learning in CAD is no longer a research topic. It already constrains sketches, drafts drawings, predicts commands, answers questions and predicts simulation results, and text-to-CAD and generative 3D models are close behind. The teams that gain most are not the ones that hand design over to AI, but the ones that use it for repetitive work, keep an engineer in charge of every decision, feed it clean data and know exactly where their design files go. For a wider view of the AI technologies behind these tools, see our overview of breakthrough AI and ML technologies transforming industries.

FAQ

What is machine learning in CAD?

Machine learning in CAD means models trained on design data, such as sketches, 3D models, drawings and simulation results, that predict something useful in the design workflow: missing constraints, drawing views, the next command, a new shape or how a design will perform.

How is AI used in CAD software today?

The most common uses are sketch auto-constraining (Autodesk Fusion AutoConstrain), automated 2D drawings, command prediction (Siemens NX adaptive UI), help assistants (Onshape AI Advisor, SOLIDWORKS AURA), drawing clean-up (BricsCAD Blockify) and physics AI that predicts simulation results (Ansys SimAI).

Is generative design the same as machine learning?

No. Generative design computes shapes from the loads, materials, constraints and manufacturing methods you define. Machine learning learns from past data and predicts. Modern tools increasingly combine the two.

Can AI create a CAD model from text?

Yes, at an early stage. Tools such as Zoo Design Studio create editable 3D models from a prompt, and research models such as Autodesk's Project Bernini generate shapes from text or images. Results are best treated as concepts that an engineer refines.

How much data does physics AI need?

It depends on the tool and the problem. Ansys says SimAI typically needs 30 to 100 simulation results to produce a model with sufficient accuracy, and it gives each prediction a confidence level.

Will machine learning replace CAD engineers?

Not on current evidence. Today's ML features remove repetitive work and speed up exploration, but design intent, judgement, standards compliance and sign-off stay with engineers.

Is it safe to upload CAD files to AI tools?

Only after you check the terms: where the data is processed, whether it trains shared models, whether you can opt out and who owns the output. Keep sensitive models under access control and encryption.

Who owns a design created with AI help?

It depends on the jurisdiction, the tool's terms and how much a human contributed. Keep records of your own design work and register dated proof of your designs so you can show what you created.

What datasets are used to train machine learning for CAD?

Public research datasets include the ABC dataset of one million CAD models from Onshape and Autodesk's Fusion 360 Gallery dataset of construction sequences, labelled parts and assemblies. Vendors also train on their own data under their own terms.