Obtaining an initial model for acoustic full waveform inversion using generalized regression neural networks

Abstract

Acoustic full waveform inversion is a featured method extensively used to obtain subsurface velocity models. The acoustic full waveform inversion approximation based on the derivative method has the limitation of being trapped in local minima. To overcome this problem, an initial velocity model in the vicinity of the global minimum should be used as the starting point. Artificial neural networks can be used to build such initial models. In this study, a generalized regression neural network approach was applied to overcome this problem. The test results on the Marmousi and SEAM synthetic data demonstrate that the initial model estimated with the generalized regression neural network provides a better starting point for full waveform inversion. In addition, the number of iterations required to search for optimal results was reduced significantly. The reduction in the number of iterations due to determining an initial model with generalized regression neural networks also substantially reduced the computational time and reduced the probability of the model becoming stuck in local minima. The acoustic full waveform inversion yields a detailed velocity model when using the initial velocity model produced by general regression neural networks.