Advanced corrective training strategy for surrogating complex hysteretic behavior

Abstract

Despite the advances in modeling component behavior, high-fidelity finite element simulation remains challenging and is limited by the computational efficiency of large-scale models. A predictive simulation framework based on deep learning is proposed to provide accurate and efficient hysteretic models for structural analysis. The proposed framework is based on the Transformer network architecture and enriched by an advanced corrective training (ACT) strategy. The ACT strategy discovers and activates the corrective ability of the model and significantly improves the prediction accuracy for complex load cases. In addition, a compound prediction strategy is proposed. By incorporating an integrated assessment of absolute and relative losses, the proposed framework provides a sophisticated prediction capability that is profoundly beneficial to global simulation. Its superior accuracy and efficiency are illustrated through case studies on both a refined steel brace model and the Bouc-Wen hysteretic model. Finally, a data-physics coupling driven structural simulation method is developed. The efficiency of the proposed method is two to three orders of magnitude higher than the classical finite element analysis, while the accuracy is almost identical.