An adaptive physics-informed deep learning approach for structural nonlinear response prediction

To effectively respond to severe seismic events, accurate and efficient models for predicting structural performance are essential. In this study, a multi-task adaptive learning framework that integrates physical information from nonlinear hysteretic models into a Long Short-Term Memory (LSTM) network is proposed. This framework overcomes the shortcomings of purely data-driven approaches, which often overlook physical laws, by enhancing computational efficiency through the integration of gradient interactions, multi-task gradient loss factors, and dynamic balance weight strategies. Benchmark tests demonstrate that our method outperforms conventional physics-informed neural networks, achieving higher accuracy in earthquake response predictions.