The GRACE satellites have been orbiting the Earth since 2002, monitoring the time variable gravity field. Some of the observed fluctuations are due to geodynamic causes, but they are often hidden in the complex signal, composed of hydrology, ocean, atmosphere, and geodynamics, the signal of geodynamic origin being usually the smallest. In addition, dealiasing residuals and noise make the separation of the signal from the different causes more difficult. We proposed a method based on the Empirical Orthogonal Function decomposition to extract the signal of physical origin, under the hypothesis that the physical signal is spatially more consistent than the noise and aliasing incomplete correction. We used synthetic geoid variations associated with earthquakes located at nearly 2000 positions at the Earth surface, based on several examples of large actual subduction events. We show that, with the present day accuracy, we can retrieve the geoid variations associated with more than 98 per cent of the earthquakes of magnitude 9 or above, around 60 per cent for magnitude 8.8, 40 per cent for magnitude 8.6 and 33 per cent for magnitude 8.3. Some events, with the right properties and location, can be detected with magnitude as low as 8. We then applied the method to the GRACE solutions, and retrieved the Hokkaido event (2003) and the Sumatra event (2004), which is in agreement with the retrieval rates mentioned here above.