Machine learning-enabled inverse design of heterogeneous impact-resistant metamaterials

Designing impact-resistant responses in heterogeneous multi-material metamaterials remains challenging because the discrete material-layout space is high-dimensional and non-intuitive, particularly when specific energy absorption (SEA) must be increased without amplifying the peak impact force in applications such as unmanned aerial vehicle (UAV) airdrop. Despite recent advances in inverse design, data-driven optimization of impact response remains constrained by the scarcity of high-fidelity datasets that jointly represent dynamic loading and material failure. This study develops a machine-learning-enabled inverse-design framework for non-periodic heterogeneous multi-material metamaterials using a fracture-aware finite-element dataset generated under quasi-static compression and dynamic impact. A ductile damage model is used to represent polylactic acid (PLA) failure, and a Transformer-based surrogate learns the mapping from cell-wise PLA/thermoplastic polyurethane (TPU) layouts to the stress–strain response and peak impact force. The surrogate is coupled with the non-dominated sorting genetic algorithm II (NSGA-II) to identify Pareto-optimal designs in the combinatorial layout space. Across comparable designs, the optimized layouts achieved up to 58% higher SEA and 47.4% lower peak force, and the predicted improvements were verified by quasi-static compression and drop-weight impact tests. In UAV airdrop experiments, the optimized design produced a payload-acceleration peak comparable to that of pure TPU while avoiding the pronounced rebound and overturning observed for the compliant single-material box. These results demonstrate that spatially heterogeneous multi-material layouts can balance energy absorption, force attenuation, and post-impact stability.