Machine learning-based methodologies for probabilistic prediction of random seismic frame structural response

This paper proposes an innovative methodology that synergistically combines machine learning techniques with probabilistic learning on manifolds for generating samples to predict the response distribution of frame structures. Through a rigorous feature engineering process, 11 seismic feature parameters and one structural feature parameter were judiciously selected. Leveraging a small-scale dataset, an exhaustive model selection process was undertaken, evaluating the performance of Support Vector Regression, Random Forest, and Gradient Boosting Trees, ultimately identifying the optimal machine learning model. By concurrently accounting for the stochastic nature of seismic motions and structural characteristics, this methodology is employed to predict the distribution of structural responses of multi-story reinforced concrete frame structures subjected to stochastic seismic events. The results demonstrate that this methodology achieves a high degree of prediction accuracy on the test dataset and can reasonably predict the seismic damage to reinforced concrete frame structures, thereby providing valuable guidance for post-earthquake disaster assessment and emergency response efforts.