Coupled data/physics-driven framework for accurate and efficient structural response simulation
Guanghao Zhai, Billie F. Spencer, Jinhui Yan, Wenjie Liao, Donglian Gu, Carlotta Pia Contiguglia, Cristoforo Demartino, Yongjia Xu
Abstract
Achieving accurate and computational efficient simulations is vital for the design, construction, and maintenance of buildings and infrastructures. Traditional physics-driven methods, such as the finite element method , struggle to balance precision with computational efficiency. In contrast, data-driven methods, such as deep neural networks , fall short in generalization and robustness. Therefore, this study proposes a coupled data/physics-driven simulation framework to harness the advantages of data- and physics-driven models, to achieve accurate and computational-efficient structural response simulation. First, the overall concept of the proposed framework is outlined, including modeling and separating the target structure into data- and physics-driven sections. Based on the discussion of the fundamental approaches for data-driven simulation, an innovative attention-enhanced stacked regression neural network is proposed to improve the accuracy of data-driven section. This architecture integrates dataset augmentation method, stacked regression, and attention-based feature enhancement. Furthermore, physics-driven modeling and the integration between the data- and physics-driven models are investigated. Finally, a case study is conducted based on a three-story frame/shear-wall building. The results demonstrate that the proposed method achieves accuracy comparable to refined finite element models , with an average stress/strain deviation no more than 0.1 %. Meanwhile, the required computational time is similar to that of a much-simplified model, with a speed-up ratio exceeding 70 times.