Abstract

As engineering materials becomes increasingly complex, accurately predicting their mechanical behaviour under diverse loading conditions presents a significant challenge. Multiscale models offer a robust solution by bridging micro- and macroscales; however, they remain impractical due to the substantial computational demand of performing microscale computations across a macroscale domain. This study explores data-driven strategies integrated with machine learning techniques to enable the efficient homogenisation of the microscopic mechanical response of porous elastomers. A micromechanical finite element model of a porous unit cell is developed within a computational homogenisation framework to generate training data. Initially, neural network-based macroscopic surrogate models are established to predict the nonlinear elastic response of a hyperelastic porous unit cell using data from micromechanical simulations. The data-driven framework is then extended to capture the time- and path-dependent response of a viscoelastic porous unit cell. In this case, a knowledge-based data-driven approach is compared to the purely data-driven strategy and demonstrates improved efficiency while maintaining accuracy in homogenising the inelastic mechanical response.