Physics-informed machine learning for predicting specific energy absorption in additively manufactured tetra-chiral auxetic lattice structures

Abstract

Architected auxetic lattice structures exhibit high lightweight and crashworthiness potential for impact mitigation during crash events. The predictive design of architected auxetic lattice structures for energy absorption performance is often constrained by the limitations of classical density scaling theory, as it fails to accurately capture the influence of complex geometrical configurations experiencing nonlinear deformation mechanisms. In this research work, a novel physics-informed machine learning (PIML) framework is introduced which enables interpretable modeling of specific energy absorption performance of the 3D-printed tetra-chiral auxetic lattice structures during quasi-static compression test. These tests were conducted on tetra-chiral lattice specimens with varying geometrical configurations, fabricated using Fused Deposition Modeling (FDM), to evaluate the influence of ligament thickness, ring radius, and ligament length on the Peak Crushing Force (PCF), Mean Crushing Force (MCF), and Specific Energy Absorption (SEA). A density-based power-law relationship was established initially between the relative density of lattice structures and SEA. The scaling exponent of 1.2786 between density and SEA indicated combined bending and stretching deformation behavior. A PIML framework with physics-based scaling function and geometrical component was developed to predict SEA. The structured decomposition method maintains physical consistency while it enables stable learning with restricted data availability. The optimized XGBoost-PIML model achieved R² = 0.803 and SHapley Additive exPlanations (SHAP) revealed a hierarchical structure of influence which showed that the thickness-to-radius ratio as the primary geometric factor for determining SEA. From SHAP analysis, the nonlinear density–geometry coupling effects on SEA were also interpreted. The proposed hybrid framework exhibited better prediction accuracy and less overfitting with limited data availability which can be extended to other architected structures.