The 17 January 1994 Northridge earthquake (Mw 6.7) occurred on a buried thrust fault in the northwest Los Angeles metropolitan area. We investigate the source process of this earthquake using the CSMIP strong motion records and a composite source model developed by Zeng et al. (1994a) for realistic earthquake strong ground motion prediction. Our previous studies demonstrated the realism of the synthetic strong motions generated from the composite source model by comparing them with observed records from earthquakes in many areas of the world. This article addresses an inverse study of the problem to find a specific composite source model for the Northridge earthquake. This is done by adjusting the location of a suite of composite subevents, using genetic algorithms (Holland, 1975), to best match the observed waveforms. A test run of the genetic algorithm on synthetic data sets finds a very good convergence of the approach. We reduce largely the intensive computation time by identifying subevents with major contribution to the waveform fit. Our result for the 1994 Northridge earthquake indicates a complex earthquake rupture process with three large slip zones: one at the hypocenter and the other two to the west of the hypocenter. We then use this model to compute the high-frequency strong-motion velocity and acceleration. The results show that the composite source model provides a very realistic broadband source description for the Northridge earthquake.