GRAPH NEURAL NETWORKS FOR INTERFERENCE MATRIX PREDICTION IN GENERATIVE DESIGN FOR DISASSEMBLY

Generative Design for Disassembly (GDfD) is becoming increasingly critical for sustainable manufacturing, enabling efficient product disassembly for recycling and reuse. In response to the growing need for end-of-life strategies, this research introduces a novel methodology employing Graph Neural Networks (GNNs) to predict interference matrices, a crucial constraint for GDfD implementation. Interference matrices hold the spatial relations and potential collisions between individual parts of the assembly during the disassembly. The proposed method takes advantage of machine learning techniques to predict potential disassembly conflicts in the assembly, which helps in disassembling the components during recycling and remanufacturing processes. Our approach innovatively modifies the underlying graph structure representing product assemblies to explicitly encode part relationships, significantly enhancing the Graph Neural Network’s ability to accurately identify the interferences during the disassembly. The modified graph structure also helps the GNN to learn and interpret complex disassembly relationships. The evaluations using this modified graph structure demonstrate substantial improvements across key performance metrics. While these initial results are promising, further research is necessary to scale the database to encompass diverse assembly types and refine GNN architectures to capture more complex disassembly patterns. Overall, this work lays a fundamental groundwork for realizing effective GDfD solutions that promote sustainable manufacturing practices and circular economy principles.