Text-Guided CAD Model Retrieval via Graph Neural Networks and Cross-Attention

Sammanfattning

In industrial Computer-Aided Design (CAD) workflows, engineers frequently customize existing designs to meet specific requirements. Over time, this leads to accumulating repositories of CAD model variations. These repositories often have inconsistent or missing metadata, making search and retrieval difficult and limiting reusability. This thesis addresses the problem of retrieving existing CAD models based on both a base CAD assembly and a textual description of required modifications, a task referred to as Text-CAD retrieval in this thesis. A dual-encoder architecture is proposed, combining a graph-based representation of CAD assemblies extracted from STEP files together with a sentence encoder. A graph encoder captures structural relationships within CAD assemblies using a graph attention network, while the sentence encoder transforms modification descriptions into a shared embedding space. These representations are fused using a cross-attention mechanism and aligned using InfoNCE contrastive learning. The system is trained and evaluated using an industrial dataset from Fagerhult containing 596 CAD customization cases. Performance is measured using the retrieval metrics: Mean Reciprocal Rank (MRR)and Hit Rate, while the embedding space is evaluated with metrics such as Active Dimensions and Effective Rank. The best-performing model achieves an MRR@5 of 0.453 and a Hit Rate@5 of 0.648, indicating that the system can retrieve relevant prior designs in a significant portion of cases. The results demonstrate that cross-modal embedding alignment between CAD structures and textual modification descriptions is feasible, but also reveal limitations in semantic understanding for the proposed model. In particular, performance is hindered by domain-specific expressions in the modification descriptions, as well as challenges in capturing specific key attributes critical for CAD model reuse. This suggests that future improvements should focus on more effective encoding of domain-specific language and the incorporation of multiple valid target designs during training to better capture what constitutes relevant CAD attributes. Despite these limitations, the results show clear potential for learning structured relationships between design geometry and text, indicating that Text-CAD retrieval systems can become a valuable tool in industrial design workflows.