Inverse unit-cell design for truss lattices with percolation-inspired peak-stress tailoring

Metamaterials offer strong potential for lightweight impact mitigation because their architecture can be tailored to control stiffness, energy absorption, and transmitted stress. However, their practical design remains limited by the cost of generating large numerical databases and by the difficulty of solving inverse design problems, where unit-cell parameters must be identified from target macroscopic responses. This study presents an automated finite element workflow for generating a dataset of periodic truss lattices and training a Unit-cell Optimization Network (UON) for inverse unit-cell design. Kelvin and octagonal beam-based lattices are generated parametrically, simulated under dynamic compression, and post-processed to extract four mechanical descriptors: initial stiffness, plateau stress, densification strain, and absorbed energy. After cleaning, 1376 valid simulations are retained for UON training and evaluation. The inverse model combines regression for the continuous strut thickness with classification for the discrete unit-cell size, achieving a test RMSE of 0.012571 mm for thickness prediction and a cell-size classification accuracy of 92.7536%. In parallel, a percolation-inspired study is performed separately from the UON training dataset by randomly removing beams according to an assigned absence probability. The results show an approximately linear decrease in plateau stress and absorbed energy with increasing absence probability, while peak stress follows an exponential-type reduction. These findings demonstrate that automated dataset generation, mixed regression-classification inverse modelling, and controlled connectivity reduction can support efficient tailoring of lattice energy absorption performance.