Abstract: The rapid prediction of the seismic response of rocking rigid blocks such as furniture, equipment, in buildings holds significant importance. To address the time-consuming challenge of numerical analysis methods, this study proposes a machine learning-driven approach for the rapid prediction of the seismic response of rocking rigid blocks in buildings. In this method, the seismic responses of the building at different floors are calculated using city-scale nonlinear time-history analysis. The seismic responses of the rocking rigid block are calculated by inputting the seismic response of the building floor into the rocking rigid block model. Machine learning algorithms are then adopted to predict whether the rigid blocks will overturn. Taking a rocking rigid block located in a four-story frame structure as an example, the proposed method is demonstrated, and five commonly used machine learning classification algorithms are compared. The influence of different input variables on whether the object overturns is analyzed, leading to the following key conclusions: The random forest algorithm exhibits the best performance in predicting whether rocking rigid blocks in buildings will overturn; Peak ground velocity is identified as the critical variable affecting whether a rocking rigid blocks overturns, followed by cumulative absolute velocity and Arias intensity, while peak ground acceleration and significant duration have relatively minor importance, though they still influence the predictive model; The method proposed efficiently and accurately predicts whether rocking rigid blocks in buildings will overturn, achieving an overall accuracy of 95.5%, and enhances computational efficiency by 4287 times compared to numerical analysis methods, providing a valuable reference for post- earthquake emergency assessments.