Inverse Design of Programmable Mechanical Metamaterials via Explainable Graph Neural Model

ABSTRACT

The rational design of metamaterials enables anomalous physical properties, yet achieving target performance remains challenging in ultra‐large optimization spaces. Herein, we propose an intelligent design framework for programmable mechanical metamaterials based on graph neural networks (GNNs). A dataset of programmable mechanical metamaterials was first generated by finite element analysis to provide a diverse design space associating structural topology with mechanical performance. To overcome the limited global receptive field of conventional GNNs, a multi‐scale structure‐aware attention module was developed. Its effectiveness and generalizability were validated across multiple public datasets and benchmark models. By incorporating geometric constraints, the enhanced GNN model achieved high‐fidelity prediction of nonlinear stress‐strain curves with an accuracy of 97.65%. GNNExplainer‐based interpretability analysis further identified critical subgraph structures and node features driving model predictions, thereby improving the transparency and reliability of the proposed framework. In addition, a graph generation strategy based on edge‐feature was introduced to enable the inverse design of metamaterials with user‐defined relative densities and mechanical parameters. The designed architectures were jointly verified by numerical simulations and experimental validation, confirming the practical feasibility of the proposed framework. This work provides a robust and generalizable intelligent pathway for both property prediction and programmable generation of mechanical metamaterials.