Deep learning architectures for automated change detection and classification in mechanical design

• Introduces 3D-CDC: a novel approach for detecting and classifying design changes in 3D CAD models. • Presents 3D-GCD, a new dataset of geometrically altered CAD components with multi- format visual representations. • Evaluates four deep learning models for classifying engineering relevance of design changes. • Demonstrates that CRNNs achieve strong performance with fewer parameters, supporting efficient AI deployment. • Advances human-AI collaboration in CAD by enabling interpretable and workflow-integrated design change analysis. This paper addresses the challenge of supporting dynamic mechanical design processes in distributed, multi-user environments. Although mechanical CAD is a mature technology, assessing the functional and manufacturing impact of design changes remains a labour-intensive manual task. To address this issue, the authors introduce the concept of “3D CAD Change Detection and Classification” (3D-CDC) and present a novel dataset, “3D-GCD” (3D Geometric Change Dataset), consisting of pairs of 3D CAD components with a variety of technically plausible changes. The dataset represents 3D components in multiple 2D image formats (e.g., grayscale and comparative heatmaps) with renderings generated from single or multiple viewpoints. This dataset is used to train four deep learning architectures—Siamese Networks with convolutional encoders, transformer encoders, 3D Convolutional Networks (3D Conv), and Convolutional Recurrent Neural Networks (CRNN)—to classify the engineering significance of changes to a component’s design. The results indicate that while three architectures (Siamese Networks, transformer encoders, and CRNN) exhibit similar accuracy, the CRNN demonstrates greater consistency and efficiency, requiring fewer trainable parameters for 3D-CDC tasks.