Machine Learning-Driven Multiscale Modelling of Shock-Absorbing Hyper-Elastic Metamaterials
Please login to view abstract download link
This work develops computational tools for the design of shock-absorbing dissipative metamaterials, focusing on rapid shape recovery and high-frequency impact resilience. While hyper-elastic materials are typically non-dissipative, this study demonstrates that macroscopic dissipation can be achieved through the emergence of a non-convex free energy, even when using polyconvex hyper-elastic constituents at the lower scale. The microscale architecture is defined by a buckling lattice of periodic cells. A multiscale framework is established to couple the 2D/3D macroscopic solid domain with the beam-based microstructure. This approach effectively handles unmatched dimensions between scales while precluding instabilities through the use of dimension-perturbed buckling layers within the Representative Volume Element (RVE). Machine learning techniques were employed to reduce computational costs by constructing a microscale surrogate model. The model is trained using homogenized data from multiple microscale experiments. Once trained, the model rapidly predicts the microscale effective material properties. This strategy presents a new alternative for integrating machine learning techniques into the analysis of metamaterials.