These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Whilst computers can store, access, render and manipulate Computer Aided Design (CAD) models they have no understanding of what is being represented. Many processing steps in CAD are still completed manually. This is because a machine lacks the engineer'™s tacit understanding of part features and what they are for. These manual tasks can be a significant proportion of the cost in design iteration. Enabling machine feature identification can enable the automation of downstream processes like design for manufacture. Current research focuses on lower-level features, specifically the process used to make the feature, e.g. punch or subtraction. It does not examine how those features combine into a useful unit in the design. The purpose of this research is to demonstrate identification of high-level features using graph neural networks (GNN). CAD designs are generated automatically, forming a range of complex parts, with a focus on the structures seen in aerospace composites. Holes, fillets, and radii are chosen as selected features and investigated. These parts were labelled and converted into step format. After further processing, they were converted into GraphML format, which is compatible with Pytorch Geometric. The features added for nodes include face orientation, count of edge type (per the step definition), the basis of the faces, the total degree and for edges; the specific curve type of that edge. The graphs are tagged at node level with binary classes, i.e hole or not hole. These classes are imbalanced with a bias away from the class of interest. This is more prevalent in smaller features as these have fewer nodes that make them up. Models were trained to maximise against Area Under Curve Receiver Operating Characteristic (AUC RoC) and down selected to a Graph attention network architecture (GAT) over GNNs and Graph Sage. Initial results showed that the models were learning relative to the origin. New samples were generated at random positions compared to the point (0, 0, 0). The models are binary classifiers as part of a related project. The models demonstrate a high level of capability (AUC ROC ~ 99%) against the generated synthetic data. We also hand labelled a set of real CAD models for the hole feature. There was significant degradation in performance against the real CAD models (AUC ~91%). Potential factors focus on improving the variability of the generator to generate more partially over-lapping features. Work identifying additional features would also be beneficial overall. This paper demonstrates: ? A new mechanism for generating part models with composite features present in aerostructures ? GNN models show a great deal of promise on synthetic data with high accuracy and AUC RoC ? Demonstration of performance degradation against real data