Toward automatic calibration of constitutive models through data-driven inverse learning

This study introduces a novel framework for automatic constitutive model calibration in which a neural network learns parameter–response relationships directly from simulated material behavior. Statistical sensitivity analysis first quantifies the influence of each model parameter on physically interpretable response features extracted from drained monotonic, undrained monotonic, and cyclic undrained triaxial simulations, associated with stiffness, shear strength, liquefaction potential, dilatancy, and critical state characteristics. The resulting parameter-influence measures are incorporated into the inverse-learning objective as weighting factors, ensuring that prediction errors associated with more influential parameters receive greater emphasis during neural network training. Physically admissible parameter combinations are subsequently sampled, and each parameter set is used to run constitutive simulations across different initial states and loading conditions, forming a database of parameter–response histories. A bidirectional long short-term memory network (LSTM) is trained on this database to map complete stress–strain histories to the model parameters that generated them. The predicted parameters provide an informed initial estimate for a Genetic Algorithm (GA), which further refines them by directly minimizing the discrepancy between simulated and experimental responses. By learning the inverse relationship directly from constitutive model simulations, the proposed framework avoids the need to reformulate the governing equations and can therefore be applied across different constitutive formulations. Its generality and performance are demonstrated using both elastoplastic and hypoplastic models, with the resulting calibrations benchmarked against conventional manual calibration. Comparison of non-optimized and GA-optimized calibrations reveals markedly higher accuracy for the GA-optimized cases, highlighting the critical contribution of the GA to the final calibration outcome.