Circularity-based hybrid learning framework for disassembly sequence planning in early-stage design

In the early-stage design phase, disassembly sequence planning must often rely on an incomplete engineering bill of materials rather than on detailed CAD geometry. This paper proposes a circularity-based hybrid learning framework that combines a genetic algorithm, Q-learning, and a graph neural network to generate and refine feasible disassembly sequences from the physical-contact relationships between components. The method evaluates candidate sequences against four criteria: disassembly time, component quality, adjacency preservation, and structural complexity. Across two case studies, a smart switch and a modular battery pack, the framework achieved competitive performance and improved interpretability of circularity trade-offs under incomplete product information.