Hybrid Data-Driven Mechanics-Based Method for Seismic Demand Estimation of Steel Moment-Resisting Frame Buildings

This study proposes deep learning–based surrogate models for predicting the engineering demand parameters of steel moment-resisting frames. The aim is to develop a comprehensive scenario-based approach that transitions from data-driven scenarios to hybrid methods by incorporating the time history of ground motion, structural responses, pushover curves, and building characteristics. The database comprises 11,500 samples from incremental dynamic analyses of steel moment-resisting frames used to train deep learning models. The results indicate that scenario-based models can be significantly improved by using hybrid deep learning models, particularly by integrating the time history of roof acceleration with ground motion. For predictions of peak and residual story drift ratios and peak floor acceleration, considering the time history of roof acceleration, the R-factors are 0.867, 0.829, and 0.949, respectively. Additionally, the prediction of peak floor acceleration in three damage states achieved an R-value of 93.0%. Finally, to evaluate the application of the trained deep learning algorithm, three instrumented frames, 5- and 9-story finite-element models as representatives of instrumented buildings (both unseen buildings under unseen ground motion), and a real 3-story instrumented office building, are selected as case studies. The maximum prediction errors for peak floor acceleration through height, considering the time history of ground motion and structural responses, are 10%, 7%, and 18%, respectively.