点击次数：

摘要：Rapid and accurate prediction of structural seismic responses is essential for effective earthquake hazard mitigation. Traditional nonlinear time-history analysis methods, while highly accurate, incur substantial computational costs. Conversely, simplified structural models and analyses enhance computational efficiency but compromise predictive accuracy. Recent machine learning (ML) approaches have emerged as promising alternatives; however, their effectiveness often remains constrained by insufficient training data and limited generalization capability. To overcome these limitations, this study proposes a novel deep learning method that integrates physics-informed input representations with scientifical training strategies. Specifically, response diagrams in the time domain, depicting linear response histories of single-degree-of-freedom systems, are introduced as model inputs. These diagrams effectively encode both the time and frequency characteristics of ground motions, as well as efficiently represent the solutions to the equations of motion. Leveraging these physicsinformed features, several state-of-the-art deep learning architectures adapted from the image classification domain are systematically evaluated for their ability to predict nonlinear structural seismic responses. Additionally, the study investigates the influence of various optimizers and learning rate scheduling policies on model training and predictive performance, ensuring adherence to scientifically training strategies. Furthermore, a hybrid transfer-learning framework is developed, enabling effective fine-tuning of models for different structural systems using limited datasets. By combining physical insights with advanced ML techniques, the proposed approach significantly enhances computational efficiency, prediction accuracy, and generalization capability. Through its innovative incorporation of prior physics knowledge, this work offers a robust and efficient solution for rapid seismic response prediction.

是否译文：否

收录刊物：SCI