To improve the calculation efficiency of structural seismic response analysis, the segment of strong ground motion can be used as the only input because it determines the magnitude of structural responses. In this study, a deep-learning neural network for predicting ground motion duration was proposed. The criterion used in the method is that the maximum displacement of the structure remains unchanged before and after the ground motion truncation, and the method considers the influences of period elongation, high order modes, and the uncertainty in estimating the structural yield strength. The deep-learning method can provide prediction results of ground motion duration with different structural periods. Taking parameters of ground motion and structure as the input features, the deep-learning model used 80 280 samples for training and prediction, and was applied to analyze the maximum story drift ratios of 4-story and 16-story frames respectively. The results were compared with the errors obtained from the widely used methods (95% Arias duration and 75% Arias duration). Results show that the proposed method and the 95% Arias duration method both performed well for the 4-story frame, but the calculation error of the 95% Arias duration method was larger for the 16-story frame; the errors of the 75% Arias duration method for 4-story and 16-story frames were much larger compared with the proposed method. The proposed prediction method of ground motion duration based on artificial intelligence is expected to improve calculation efficiency, reduce error, and widen the application scope.