Abstract: Machine learning has been widely applied to establish ground motion models. However, due to the spatio-temporal sparsity of strong motion data, insufficient data and model selection will respectively lead to within- and between-model epistemic uncertainties. This study proposes an epistemic uncertainty quantification method based on Bootstrap and on logic tree Monte Carlo simulation, which enables the quantification of within-model epistemic uncertainty for any machine learning architecture. Traditional models and machine learning models are introduced simultaneously in the probabilistic seismic hazard analysis, and the results of seismic hazard analysis under the influence of different epistemic uncertainties are compared and studied. The research results show that the epistemic uncertainty of the machine-learning-based ground motion model has a significant impact on both the mean value and the standard deviation of the ground motion parameter values under a given ground motion level, and that this impact is more remarkable compared with that of traditional models.