Large-dimensional seismic inversion based on global optimization and autoencoder

Seismic inversion often involves nonlinear relationships between model and data and the misfit function usually has many local minima. Global optimization algorithms are well-known capable to search for the global minimum of a misfit function without requiring a good initial model. However, these algorithms can hardly work for large-dimensional cases because of the “curse of dimensionality” problem. In this paper, we mitigate this problem by introducing a neural network called autoencoder into seismic inversion and propose a new inversion method based on global optimization and autoencoder. Benefiting from the dimensionality reduction characteristics of autoencoder, in the proposed method, the original large-dimensional problem is transformed into a low-dimensional one that can be efficiently optimized by a global optimization algorithm. Preliminary numerical examples demonstrate that the proposed method can solve large-dimensional seismic inversion problem with a significant improvement in efficiency compared with conventional global optimization based method.