Graph neural network based visual programming recommender system for 3D CAD shape evaluation

Abstract

Design for Manufacturing (DfM) is becoming increasingly important in computer-aided design (CAD) due to the growing complexity of products. Automatically evaluating design rules for manufacturability from CAD dimensions and geometry is one effective approach for DfM. However, implementing automated design-rule-checks requires coding complex programs on CAD systems and the expertise to implement them, which make it difficult to adopt the approach, or implement product-specific design-rules. In this study, to mitigate this barrier, we developed a visual programming environment for the design-rule programs and a graph neural network (GNN) based recommender system that assists users to create the visual program. The proposed visual programming environment allows users to make design rules by placing and connecting shape recognition and measurement function nodes. This simplifies and improves the programming interface. Furthermore, the proposed recommender system predicts the subsequent function node that users will append, which reduces the user’s cognitive load of choosing the right function nodes. By employing proposed GNN architecture in the visual program recommender system, the input-output relationships between the function nodes and their input arguments are naturally taken into account to produce accurate recommendations. We evaluated the performance of the proposed recommender in the off-line experiment and the experiment in real-world use. The results in both experiments demonstrate that the proposed GNN based visual program recommender system can suggest the function nodes in sufficient accuracy, which contributes to improved productivity of implementing automatic CAD design-rule-checks.

Keywords

• computer aided design
• visual programming
• recommender system
• deep learning
• design for manufacturing