Just another WordPress site - Ruhr-Universität Bochum

Machine learning (ML) approaches have emerged as a powerful alternative to classical constitutive models, which rely on postulated yield surfaces and phenomenological hardening laws. These functional forms often require extensive parameter calibration and can lead to inaccuracies due to plastic anisotropy, non-unique parameters, or experimental noise. Despite offering compelling solutions to the limitations of classical constitutive models, most ML-based approaches continue to encounter challenges—such as high data requirements, separate handling of yield onset and strain hardening, or complex architectures—especially when describing cyclic plasticity. To address these challenges, this work introduces an ML-based constitutive model for cyclic plasticity, capturing isotropic, kinematic, and mixed hardening in a single classification step via support vector classification (SVC). By learning the yield surface and its evolution via a global history-dependent feature set, the model defines an associative flow rule for cyclic plasticity that captures any hardening behavior in the data without feature adjustments, delivering accurate stress–plastic-strain predictions. Training datasets, generated separately for isotropic, kinematic, and mixed hardening, span a diverse range of loading directions, enabling proper generalization. Validation across varied conditions, including unseen loading directions, different load ratios, extended cycles, and both interpolation and extrapolation cases, confirms accurate stress–plastic-strain predictions, with classification metrics indicating high accuracy, precision, and recall. The thermodynamic admissibility is ensured by the decision boundary’s outward gradient, preventing any non-physical material response under proportional cyclic loading. Overall, this classification-based approach offers a computationally efficient and interpretable alternative for data-driven plasticity modeling under cyclic loading.