Inversion of 1-D magnetotelluric data using CNN-LSTM hybrid network

The magnetotelluric (MT) inversion is nonlinear and ill-posed, which poses great challenges for accurate model reconstruction. To tackle these challenges, in this study, a novel method using a deep neural network for inversion of 1-D MT data is proposed. The proposed network structure combines a convolutional neural network (CNN) for features extraction and a long short-term memory (LSTM) network for resistivity model reconstruction. The implementation of this method consists of three phases. In the dataset acquisition phase, a random sample generation scheme is proposed to ensure a sufficient number and diversity of datasets. In the training phase, an error back-propagation scheme is adopted to update the network parameters to extract and store the complex nonlinear relationship between the models and MT responses. After that, the unknown model can be reconstructed from the MT data using the trained network. Synthetic and field data are considered to verify this method. The corresponding results show that the method proposed in this paper is computationally efficient and has high inversion precision.