Unlocking the Power of CAD Data in AI

AI tools are touching every industry, including engineering, with large language models, image-trained tools, and custom AI agents finding their way into applications across markets. But one frontier remains largely hidden: CAD-informed machine learning models for use in engineering applications.

The potential value is intriguing; tools that could recognize the relevant context of your specific CAD data, incorporating assemblies and geometry, metadata, tolerances, and more to provide holistic insights that inform every stage of engineering workflows. Making practical use of this data in AI tools is both difficult and immensely valuable.

The same richness of data that makes this such an enticing prospect makes it uniquely challenging to utilize in machine learning tools. New tools are breaking down these barriers, providing developers with the fastest path from CAD data to working machine learning models.

Why is CAD Data Such a Challenge to Use in AI Tools?

Engineering CAD data poses unique obstacles for AI developers. Unlike images or text, CAD files are:

• Highly structured, hierarchical, and interdependent - Assemblies reference parts, which reference features, which may encode tolerances, PMI, and design intent.

• Continuous and precise - Small differences in dimensions can have outsized effects on predictions. A 10.00 mm shaft in a 10.02 mm hole slides freely; swap those values and it's a press fit. A model that's "almost right" can be functionally wrong

• Multi-format - There are dozens of CAD systems, each with its own native file formats and quirks. Training on one may or may not carry over to another.

Bluntly, training machine learning models directly on CAD data is a vastly different process from training on natural images or text. Models must handle assembly structure, topology, features, metadata, and geometric precision simultaneously. A small oversight in preprocessing or encoding can propagate errors through the workflow.

For a closer look at what makes CAD data so difficult to bring into AI workflows, read Why CAD Data is the Hardest Part of Bringing AI to Manufacturing Applications.

Why Do Companies Want CAD-Trained Machine Learning Tools?

Despite the difficulty, companies continue to invest in CAD-informed AI because the potential payoff is huge. Engineering organizations hold decades of detailed design data, much of which is locked inside files that are difficult to analyze at scale. CAD-trained machine learning models offer a way to extract value from that information without reducing it to oversimplified representations. Even a basic question, such as whether a similar part already exists somewhere in the archive, is difficult to answer at that scale.

How is Tech Soft 3D responding to the massive changes AI is bringing our industry? Learn more in our blog Tech Soft 3D & AI: Our Role, Our Vision, and Unlocking CAD for AI

Improving team efficiency is a clear benefit of these tools. Engineering teams spend large amounts of time reviewing models, validating intent, and checking relationships already encoded in CAD. AI tools trained on CAD data can interpret surface features, tolerances, and structural relationships far faster than manual review, allowing engineers to focus on design decisions rather than inspections.

CAD data is even more important than most realize in modern organizations. Learn more in 3 Ways Your CAD Data Is More Valuable Than You Thought.

By identifying inconsistencies earlier, CAD-trained machine learning tools can also reduce downstream manufacturing defects and rework. Once trained, these models can also extend their utility to simulation, manufacturing, quality assurance, and other engineering systems, turning accumulated design knowledge into insight that compounds over time, rather than remaining siloed in individual tools.

The specific use cases are significant, from increased automation in robotics, manufacturing, CAE, scan data, and design workflows.

How are Developers Integrating CAD into AI Models?

As with any machine learning process, data selection and optimization are key to the process. CAD data is often sensitive intellectual property. Datasets are rarely public, standardized, or immediately ready to start training a model, unlike image or text datasets. Internal archives that most organizations have were rarely created with machine learning in mind. Files span multiple CAD systems, modeling conventions, and levels of quality. Before a single model can be trained, developers must first decide what "usable" CAD data even means for their application.

Once data is collected, it must be transformed into a form machine learning algorithms can consume, and this is where many efforts stall. CAD files are not flat objects. They encode relationships between faces, edges, features, and assemblies. Choosing how to represent and encode that information for machine learning is a hard decision; some teams experiment with mesh-based approaches, others with boundary representation (B-Rep) graphs, and many discover that preserving engineering intent requires hybrid strategies. Each choice affects ML model performance, memory usage, and training time.

Training the model introduces another layer of complexity. Processing CAD-based datasets is far more computationally intensive than training with text or image data. Geometry-heavy representations increase memory requirements, while batch sizes must be carefully managed to avoid overfitting to specific design patterns. As AI algorithms evolve, ML models must evolve alongside them. For developers, this turns model training into an ongoing engineering effort rather than a one-time task.

Finally, there is a question of generalization and usage. A CAD-informed machine learning model must extrapolate beyond the specific geometry it was trained on. This is where approaches such as graph-based learning become critical, allowing models to infer relationships with unseen assemblies rather than memorize shapes. Even so, validation remains difficult. Developers must test against real-world designs, often discovering edge cases that force them back to earlier stages of the pipeline.

Taken together, this workflow explains why CAD-informed AI remains in its infancy. It demands fluency across CAD systems, machine learning theory, and software infrastructure. Perhaps most limiting, it takes a special set of data and computer science skills to execute efficiently.

The effects of AI on engineering and design spaces are difficult to overstate. Learn more from one of our recaps on How AI is Impacting Engineering, Design and Simulation

Naturally, this leads to the question: how can companies meaningfully reduce the resource cost of CAD-based machine learning to take advantage of the huge potential the technology offers?

From Challenge to Capability: Enter HOOPS AI

Tech Soft 3D has spent over 30 years providing toolkits for CAD data access, visualization, and CAE applications, handling more CAD data than just about anyone in the industry, and HOOPS AI brings that experience to machine learning. The framework is built for machine learning engineers and data scientists training models on their own CAD data.

The HOOPS AI software development toolkit provides the pipeline components teams would otherwise build themselves, including data access, ingestion, versioning, lineage, and reproducible runs. HOOPS AI does not deliver pre-built answers to industry problems and will not replace your engineers. Teams still need curated datasets and domain knowledge, and most industrial applications will require custom training.

Tech Soft 3D partners are already leveraging HOOPS AI alongside other HOOPS toolkits. Learn more about how Synera uses HOOPS AI in our Synera & Tech Soft 3D partner success story.

What does HOOPS AI Do?

HOOPS AI builds on HOOPS Exchange to read a wide range of CAD formats through a consistent API. It supports B-Rep and hybrid encodings, graph-based assembly structures, and manufacturing annotations, allowing models to learn from topology and relationships rather than simplified shapes. Sample models cover part classification, feature recognition, part similarity search, and feature similarity matching. Training runs on either CPU or GPU, or you can use pre-trained models based on publicly available CAD data. Teams can integrate their own encoders and preprocessing steps and deploy in private or cloud environments, so data ownership stays with the organization.

Finding Similar Parts with Shape Embeddings

For many engineering organizations, the first useful application of CAD-informed machine learning is answering the question raised earlier: whether they have already designed something like the part in front of them. HOOPS AI uses graph neural networks on B-Rep data to convert each part into an embedding, a vector in which geometrically similar parts sit close together. Developers can embed a library of CAD files, index it, and query it with a CAD file to return the closest matches. Indexes support metadata filtering and can be saved and reloaded, so a CAD library does not need to be re-encoded each session. This is the basis for duplicate detection and design reuse across large part libraries.

HOOPS AI also includes pre-trained embedding models based on public datasets such as ABC (~1 million parts), Fabwave and the SIGNAL architecture (~two million parts), along with a reranking step that refines the order of the top results. For the best results on an organization's own parts, Tech Soft 3D recommends training a custom model on their own parts, which then runs through the same search interface.

The Context Layer of HOOPS AI extends search results into predictions. Using metadata the developer supplies for the indexed parts, it estimates attributes for a query part, such as material, manufacturing process, cost, or lead time. Each estimate can be tagged as ready to propose, needs review, or insufficient evidence, so applications can send low-confidence results to an engineer instead of presenting every answer with equal weight.

To see similarity search in action, watch the Understanding Shape Intelligence segment of our HOOPS AI webinar.

For developers and those who use HOOPS Exchange & HOOPS Visualize for Desktop, we have a forum showcasing how HOOPS AI can work at scale, called: HOOPS AI Inside a HOOPS Visualize Desktop + Exchange Application: Similar-Part Search Across 10,000 CAD Files

Looking Forward to CAD-Informed AI

CAD-informed AI represents a frontier that is only beginning to take shape. As AI becomes more pervasive across engineering workflows, the ability to train models directly on CAD data will increasingly define what is possible, separating surface-level automation from systems that can work with the structure and relationships inside engineering data.

HOOPS AI is one part of that emerging landscape, making the problems of training machine learning models more approachable without pretending they're solved. The tooling and the underlying techniques continue to be shaped by developers, researchers, and companies willing to invest early.

The organizations engaging with CAD-informed AI today are helping define how engineering knowledge is captured, reused, and amplified in the years ahead.

To try HOOPS AI on your own CAD data, visit the HOOPS AI product page to request an evaluation, or contact us directly.

CAD Data In AI FAQs