Two-dimensional deep learning inversion of magnetotelluric sounding data

Abstract Conventional linear iterative methods for magnetotelluric sounding (MT) suffer from considerable limitations related to difficulties in selecting the initial model and local optima. On other hand, conventional intelligent nonlinear exhibit slow convergence low accuracy. In this study, we propose an inversion strategy based on deep learning (DL) belief network (DBN) realise instantaneous of MT data. A scaled momentum rate is introduced improve performance restricted Boltzmann machine during DBN pre-training stage, a novel activation function (DSoft) enhance global optimisation capability fine-tuning stage. To address difficulty designing sample data when prior information lacking, employ k-means++ algorithm cluster field use clustering results as guide construction dataset. Then, proposed DBN, ensure end-to-end mapping directly apparent resistivity model. We implement two groups experiments demonstrate validity both improvements DBN. consider six types geoelectric test set feasibility effectiveness method 2D inversion, which further compare with well-known least-square regularisation several extended models The qualitative quantitative analyses show that DL promising it can accurately delineate subsurface structures perform rapid inversion.