A Theory-Guided Encoder-Decoder model for short- and Long-Horizon seismic response prediction of nonlinear Single-Degree-of-Freedom systems

In the domain of structural engineering, accurate real-time prediction of structural dynamics is of paramount importance. In recent years, there has been a notable increase in the utilization of deep learning methodologies for the estimation of peak and comprehensive seismic responses. The success of deep learning-based surrogates in previous studies has highlighted their potential in replicating the dynamic behavior of their target structure, thereby enabling the decoding of the input excitations to their corresponding structural responses. However, the utilization of surrogates for non-specific structural systems remains under-explored, underscoring the necessity for further model adaptation prior to its deployment for different structural configurations. Building on these observations, this research presents an innovative approach that enables the surrogate model to autonomously learn the decoding mechanism from structural parameters, hysteresis curves, and lookback excitation-response data, while restricting the outputs with a modified hard constraint projection scheme. These modifications render the surrogate applicable to robust long-horizon response prediction for structural systems with nonspecific structural configurations. Comprehensive testing under various ground motion scenarios and distinct structural setups has yielded consistent predictions compared to numerical simulations, thereby validating the efficacy and adaptability of the proposed encoder-decoder surrogate model in accurately forecasting seismic responses across diverse contexts.