[Paper Review] Geometric Deep Learning for Computer-Aided Design: A Survey

Negar Heidari, Alexandros Iosifidis|arXiv (Cornell University)|Feb 27, 2024

Advanced Numerical Analysis Techniques4 citations

TL;DR

This survey presents a comprehensive review of geometric deep learning (GDL) methods for computer-aided design (CAD), focusing on learning from boundary representation (B-Rep) data to enable automated design tasks such as similarity retrieval, 3D model generation, and CAD reconstruction. It introduces key GDL architectures like graph neural networks and Transformers, benchmarks datasets, open-source tools, and identifies challenges in reproducibility and data scarcity, proposing future research directions in B-Rep generation and cross-format synthesis.

ABSTRACT

Geometric Deep Learning techniques have become a transformative force in the field of Computer-Aided Design (CAD), and have the potential to revolutionize how designers and engineers approach and enhance the design process. By harnessing the power of machine learning-based methods, CAD designers can optimize their workflows, save time and effort while making better informed decisions, and create designs that are both innovative and practical. The ability to process the CAD designs represented by geometric data and to analyze their encoded features enables the identification of similarities among diverse CAD models, the proposition of alternative designs and enhancements, and even the generation of novel design alternatives. This survey offers a comprehensive overview of learning-based methods in computer-aided design across various categories, including similarity analysis and retrieval, 2D and 3D CAD model synthesis, and CAD generation from point clouds, and single/multi-view images. Additionally, it provides a complete list of benchmark datasets and their characteristics, along with open-source codes that have propelled research in this domain. The final discussion delves into the challenges prevalent in this field, followed by potential future research directions in this rapidly evolving field.

Motivation & Objective

• To provide a comprehensive survey of geometric deep learning techniques applied to CAD data, especially boundary representation (B-Rep) formats.
• To identify and analyze the limitations of existing machine learning methods in capturing geometric features from parametric CAD models.
• To compile and review benchmark datasets and open-source implementations to support reproducibility and progress in the field.
• To highlight key challenges such as data scarcity, lack of paired datasets, and dependency on proprietary CAD kernels.
• To outline future research directions, including generative modeling of CAD operations and cross-format synthesis from point clouds or sketches.

Proposed method

• The survey employs a systematic categorization of GDL methods into two main areas: CAD representation learning (classification, segmentation, retrieval) and CAD generation (2D/3D construction, command sequence modeling).
• It evaluates supervised and self-supervised GDL models, including Graph Neural Networks (GNNs) and Transformers, trained directly on B-Rep data to preserve geometric and topological structure.
• The paper analyzes methods that bypass mesh conversion by learning directly from B-Rep, avoiding information loss and computational overhead.
• It reviews techniques for generating CAD models from point clouds, sketches, and command sequences, including approaches like SolidGen and CADParser that synthesize B-Rep directly.
• The survey compiles and characterizes public benchmark datasets and open-source codebases to support method comparison and reproducibility.
• It discusses architectural choices and implementation dependencies, particularly the reliance on proprietary CAD kernels like Parasolid, which hinder academic reproducibility.

Experimental results

Research questions

• RQ1 How can geometric deep learning models effectively learn and represent geometric features directly from B-Rep data without converting to meshes or voxels?
• RQ2 What are the current limitations in benchmarking and comparing GDL-based CAD methods due to fragmented and non-standardized datasets?
• RQ3 How can generative models be extended to produce complex CAD operations such as fillets and chamfers from command sequences or 3D inputs?
• RQ4 What are the key challenges in reproducing results due to closed-source CAD kernels and lack of preprocessing documentation?
• RQ5 How can future research bridge the gap between 2D sketches, point clouds, and parametric B-Rep models in a unified, differentiable framework?

Key findings

• Geometric deep learning models, particularly GNNs and Transformers, show strong potential in learning geometric features directly from B-Rep data, outperforming traditional methods that rely on physical or mesh-based features.
• Methods that bypass mesh conversion preserve geometric fidelity and reduce computational cost, making them more suitable for industrial CAD workflows.
• Despite progress, most GDL-based CAD methods are evaluated on small, task-specific subsets of datasets, limiting generalizability and fair comparison.
• Reproducibility is significantly hampered by the lack of open-source CAD kernels and incomplete documentation of preprocessing pipelines and code dependencies.
• The absence of paired datasets—such as sketches paired with corresponding B-Rep models—limits the training of supervised generative models for sketch-to-CAD translation.
• Open-source tools based on OpenCASCADE are more accessible, but many methods remain dependent on proprietary kernels like Parasolid, restricting academic access and innovation.