Machine Learning Accelerates Programmable Mechanics in Isotropic Diamond Plate Lattices

Abstract

This study presents a cascaded, integrated machine-learning framework for the rapid design and optimization of fully isotropic, diamond-inspired composite plate-lattice metamaterials (DMSCs), extendable to any composite architecture. An artificial neural network (ANN) trained on finite-element simulation data quickly screens large design spaces for isotropy, achieving a roughly 10⁵-fold speed-up in candidate identification compared to conventional methods. The ANN-driven screening identifies 483 perfectly isotropic configurations, which are then used to train a Bayesian-optimized XGBoost regression surrogate to accurately predict three static mechanical metrics. Coupling these surrogates with the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) yields 209 Pareto-optimal plate-lattice designs that balance stiffness, plateau stress, and energy absorption. The top design (unit-cell width: 11.3 mm; wall thickness: 1.19 mm; scaling factor α = 0.77) was validated by high-fidelity finite-element simulations, with all predicted metrics deviating by less than 4%. Our analysis reveals that the composition scaling factor α influences mechanical performance over three times more than relative density, overturning the traditional density-centric design paradigm. Additionally, our method extends the viable fully isotropic relative density range by approximately 17%, providing a robust foundation for developing ultralight, high-strength metamaterials for demanding aerospace and defense applications.