Real-Time Regional Seismic Response Prediction of Buildings Using City-Scale Nonlinear Time History Analysis and Deep Learning

Rapid prediction of regional building seismic response is essential for post-earthquake emergency actions. This study proposes a real-time prediction method integrating city-scale nonlinear time history analysis (NLTHA) with deep learning. A city-scale NLTHA-generated training dataset was constructed by integrating ground-motion intensity measures, building attributes, and seismic responses. A multi-head attention deep neural network is then developed to efficiently predict seismic responses. Validation with 327 seismic records from 47 buildings achieved 90.83% accuracy in peak top-floor acceleration (PTFA) classification and improved computational efficiency by about 1,588 times compared with traditional time-history analysis, enabling real-time assessment for seismic emergency decision-making.