Three dimensional inversion of mise-a-la-masse-based electric data using U-Net convolutional neural network

Abstract Nonlinear inversion algorithms are away ill-pose for the limited and special data measurement, such as mise-a-la-masse-based electric method and borehole to surface electromagnetic method. It is natural to develop new inversion algorithms without calculating the the Jacobian matrix and its inverse matrix, especially for three dimensional inversion. Basing on U-Net fully convolutional neural network, a three dimensional inversion of mise-a-la-masse-based electric data is proposed in this paper. The random rotated spheroids are used to synthesize diverse geological models. A typical U-Net learning network is constructed with incorporating the Dice coefficient and the input data feedback to the trained network. Synthetic datasets testing validates the effectiveness of this proposed algorithm for mise-a-la-masse-based electric data.