Modular machine learning-based elastoplasticity: Generalization in the context of limited data

The development of highly accurate constitutive models for materials that undergo path-dependent processes continues to be a complex challenge in computational solid mechanics. Challenges arise both considering the appropriate model assumptions and from viewpoint data availability, verification, validation. Recently, data-driven modeling approaches have been proposed aim establish stress-evolution laws avoid user-chosen functional forms by relying on machine learning representations algorithms. However, these not only require significant amount but also need probes full stress space with variety loading paths. Furthermore, they rarely enforce all necessary thermodynamic principles as hard constraints. Hence, are particular suitable low-data or limited-data regimes, where first arises cost obtaining latter experimental limitations labeled data, which is commonly case engineering applications. In this work, we discuss hybrid framework can work variable modularity elastoplasticity formulation each component chosen either classical phenomenological depending available information complexity response. method tested synthetic uniaxial coming simulations well cyclic structural materials. discovered material found interpolate allow extrapolation thermodynamically consistent manner far outside domain training data. This ability extrapolate limited was main reason early continued success shortcoming learning-enabled approaches. Training aspects details implementation into Finite Element discussed analyzed.