ERI: A Foundational Framework for Explainable Assembly Topology in Disassembly Sequence

This Engineering Research Initiation (ERI) award supports foundational research to establish a framework that formally connects a product's topological design features with its emergent disassembly characteristics. Maximizing value recovery from products at end-of-use is a national priority in the United States, as it advances a resilient and economically robust circular economy by expanding domestic value retention across repair, refurbishment, and remanufacturing sectors. Central to this transition is the process of disassembly. However, a significant value-recovery gap persists: current Artificial Intelligence disassembly models are limited to prediction. These models identify optimal disassembly paths but do not provide causal explanations for how specific design features constrain disassembly, thereby disrupting the feedback loop necessary for upstream engineering improvements. This project addresses the gap between prediction and explanation, transforming design-for-disassembly into a quantitative and explanatory discipline to systematically advance high-value sectors such as heavy machinery, automotive dismantling, and aircraft disassembly. In parallel, the project will promote future scientific discovery by disseminating results through an open-source ecosystem, enabling the global research community to utilize, extend, and build upon the developed datasets and frameworks. The project examines how graph-based representation methods can interpret assembly topology to identify and explain the features that determine disassembly constraints. The research comprises three integrated tasks. First, an automated pipeline will convert 3D CAD data into multi-layered representations, optimizing interference matrix generation through swept-volume culling to remove redundant collision checks and formalizing a two-level parametric assembly graph. Second, a design-based Graph Neural Network (GNN) incorporating Graph Attention Network (GAT) and global Graph Transformer layers will encode long-range physical dependencies, enabling the model to learn disassembly sequence networks directly from CAD models. Third, a post hoc Explainable AI (XAI) framework will use Structural Causal Models and computational counterfactuals to generate causal explanations. This framework will map mathematical insights directly onto CAD geometry, providing human-interpretable results and actionable design recommendations. The overall framework will be validated computationally and through qualitative user studies with industrial partners to ensure that the derived design principles are both geometrically justified and applicable in professional practice. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.