Learning elastoplasticity with implicit layers

This work presents a novel machine learning approach for learning elastoplasticity directly from stress-strain data. Data-driven plasticity modeling is challenging due to the non-smooth transitions induced by the yield criterion and the complex nature of multi-dimensional yield surfaces. To address these difficulties, we propose an implicit layer-based architecture that formulates the elastoplastic constitutive update as a convex optimization problem with learnable parameters, representing general plastic yield surfaces as parametrized convex sets. The proposed approach exhibits strong generalization capabilities due to its embedded convex structure, achieving high performance even with limited data and in the presence of noise.