Modelling Metal Plasticity and Damage with Constitutive Artificial Neural Networks

Abstract

This study presents the application of constitutive artificial neural networks (CANNs) to model the flow stress and failure strain of steels under deformation, aiming to overcome key limitations of traditional constitutive models, such as the Johnson–Cook (JC) formulation. Two literature-based datasets are employed to train the CANNs: T24 steel for modeling the plastic flow stress and E250 steel for failure strain prediction. The results demonstrate substantial gains in predictive accuracy, with the CANN approach achieving a 75% reduction in root mean square error for flow stress and a 60% reduction for failure strain compared to the JC model. Beyond enhanced accuracy, this work highlights the flexibility of CANNs for future extensions, including the incorporation of additional input variables (i.e., Lode angle) and the modeling of damage factors designed for flow stress softening. These findings further support the potential of data-driven constitutive modeling as a robust alternative to conventional constitutive formulations and fitting in computational mechanics.