Magnetotelluric Data Inversion Based on Deep Learning With the Self-Attention Mechanism

The magnetotelluric (MT) method plays an important role in the exploration of oil and gas, geothermal resources, and deep structures due to its large exploration depth and low cost. MT data inversion is a key step in data interpretation. The geophysical inversion based on a deep learning (DL) method does not need to calculate the sensitivity matrix, and the inversion results do not depend on the initial model, which has also received much attention. The feature extraction capability of a conventional convolutional neural network (CNN) is limited. The self-attention mechanism can select the data information with relatively large contribution to the result from a lot of input information and improve the ability of key feature extraction. We present a 2-D MT CNN inversion study based on the self-attention mechanism. We first use a blocky 1-D MT inversion to quickly build the representative sample model, and then we perform a 2-D MT forward algorithm by finite element method to build the training set samples. We introduce the self-attention mechanism into the CNN and place the self-attention mechanism behind the convolutional layer. After the data are extracted from the convolutional layer, feature extraction is performed again to improve the reliability and efficiency of the prediction. We tested the algorithm using different geoelectric models and field data from the Dayangshu Basin. Our results show that the inversion accuracy is improved, especially the inversion results of field data are consistent with borehole data.