A stochastic function model of seismic ground motions is presented in this paper. It is derived from the consideration of physical mechanisms of seismic ground motions. The model includes the randomness inherent in the seismic source, propagation path and local site. For logical selection of the seismic acceleration records, a cluster analysis method is employed. Statistical distributions of the random parameters associated with the proposed model are identified using the selected data. Superposition method of narrow-band wave groups is then adopted to simulate non-stationary seismic ground motions. In order to verify the feasibility of the proposed model, comparative studies of time histories and response spectra of the simulated seismic accelerations against those of the recorded seismic accelerations are carried out. Their probability density functions, moreover, are readily investigated by virtue of the probability density evolution method.