Deep learning audio-magnetotelluric and transient electromagnetic joint inversion

Joint inversion plays a crucial role in the comprehensive interpretation of geophysical data. Recently, the rapid development of data driven deep learning (DL) techniques has brought a new perspective for geophysical inversion. Hence, this work proposes a DL based joint inversion method for Audio Magnetotelluric (AMT) and Transient Electromagnetic (TEM) data. First, we generate a set of synthetic layered resistivity models and then perform AMT and TEM forward simulations to create a sample dataset for network training. After, an end to end DL. joint inversion model, named JoATInvNet, is designed and constructed based on the well known residual network and UNet. Next, considering the difference between AMT and TEM data, two individual data normalization methods are introduced to standardize the network input data. Finally, we demonstrate the proposed JoATInvNet joint inversion method on both synthetic and field data. Compared to AMT or TEM DL inversion, JoATInvNet can effectively integrates the valid information within AMT and TEM data. It possesses lower inversion misfit, more robust anti noise ability, and more accurate and comprehensive interpretation for subsurface structures.