Mechanics-informed neural networks for modeling constitutive relation for nonlinear elastoplastic materials

Identifying constitutive relations for materials with complex behaviors remains a persistent challenge in computational mechanics. Unlike metals, granular materials exhibit pressure hardening, where increasing hydrostatic pressure enhances both shear strength and stiffness through reinforced interparticle contact forces and frictional resistance. For nonlinear elastoplastic granular materials, traditional approaches rely on empirical constitutive models derived from extensive experimental datasets, but they lack flexibility and dependent heavily on parameter calibration. This study proposes a mechanics-informed neural network (MINN) framework, leveraging physics-informed learning principles, to identify nonlinear constitutive relations for geotechnical granular materials under diverse deformation path. By embedding the second-order work criterion and enforcing time consistency for path-dependent responses, MINN significantly outperforms traditional neural networks in robustness, particularly for materials with complex loading history. By integrating finite element solvers, numerical cases further validates the framework’s efficacy, demonstrating close alignment between numerical predictions and experimental data. The dual capabilities of MINN in balancing physical constraints and data-driven adaptability enhance its versatility in elastic–plastic constitutive modeling.