Deep learning based one-dimensional inversion of magnetotelluric data, and an application in the southwestern Athabasca Basin, Canada

Abstract

The magnetotelluric (MT) method uses the earth’s natural electromagnetic signals to image the resistivity distribution of the subsurface. Its application is widespread in studies to improve our understanding of the lithospheric architecture, earthquake zones, and for the exploration of natural resources, such as critical minerals and geothermal energy. Inversion of MT data is a fundamental step in the data analysis routine to retrieve a subsurface geo-electrical model that can be used to inform geological interpretations. In order to reduce the effect of non-uniqueness and the local minimum trapping problems of conventional deterministic inversion, a data-driven mathematical method with a deep neural network can be used to estimate the subsurface properties in many geophysical inversion techniques. In this study, we investigate a deep learning (DL) inversion method composed of a multi-head convolutional neural network (CNN) architecture for 1-D MT data analysis. We created synthetic datasets consisting of 100,000 random samples of resistivity layers to train the CNN model parameters. The trained model was validated with independent synthetic datasets, and the predicted resistivity distribution displayed an acceptable resolution and reliability, which demonstrate the potential application of DL inversion for MT data. The trained CNN model was used to analyze real MT data collected in the southwestern Athabasca Basin, Canada. Predicted results from the DL method displayed a similar subsurface resistivity distribution with traditional iterative inversion. Because this approach can predict a resistivity model without multiple forward modelling operations after the trained CNN model is created, this framework is suitable to speed up multidimensional MT inversion for predicting subsurface resistivity.