Vertical Seismic Profile Inversion With Genetic Algorithms

Abstract

We describe a new genetic-algorithm (GA) inversion technique and apply it to a vertical seismic profile (VSP) inversion problem where the goal is to recover slowness and impedance profiles. Our algorithm consists of one long (300 generations) and a series of short (50 generations) GA runs. After each run the problem is reparametrized using the matrix of partial second derivatives in a neighbourhood of the best model found in that run. The new unknown parameters are more nearly independent than the old ones, thus successive GA runs find better models faster, in accordance with the principle of meaningful building blocks. For the VSP problem we use a fitness function that is the weighted sum of the cross-correlation between the measured and synthetic total wavefields and the cross-correlation between the measured and synthetic upgoing wavefields. The measured upgoing wavefield is obtained from the measured total wavefield by f-k filtering.