Abstract: Ground motion synthesis methods provide input time histories for structural seismic design and performance evaluation. The existing synthesis methods can only match the target spectrum to consider the frequency spectrum and amplitude of the ground motion, but cannot consider the duration characteristics. To synthesize ground motions which matches the three elements of ground motions for a specific area, a method that incorporates machine learning is proposed. In this method, the data-driven principal component analysis is used to extract the characteristic mother waves from the actual ground motion database of the target area. The target response spectrum and duration are obtained by the ground motion prediction equation of the region. The multi-objective genetic algorithm is applied to solve the linear combination coefficient of the mother waves, so that the combined ground motion can match the preset error standard and the multi-objective parameters of the regional ground motion. The method is verified for the 2019 M_s 6.0 Changning earthquake in Sichuan Province. The results show that, because the proposed method is driven by the actual ground motion data in the target area, the synthetic ground motion matches the demand of the target spectrum and the duration characteristics of regional ground motions, which can provide more reasonable ground motion inputs for seismic analyses considering regional seismic characteristics.