ADRN: Attention-based deep residual network for Magnetotellurics (MT) inversion

Traditional linearized inversion methods are often limited by the initial model selection in complex geological scenarios, leading to inaccurate inversion results due to being trapped in local extremes. In contrast, deep learning algorithms offer powerful nonlinear fitting capabilities and show promise in electromagnetic data inversion. This paper presents the attention-based deep residual network (ADRN) for magnetotellurics (MT) inversion. The ADRN utilizes deep residual networks and attention mechanisms to improve the accuracy and efficiency of MT inversion. To prevent overly complex geoelectric models that could impact training, a method is proposed to generate a dataset by introducing control layers and adjusting their locations and resistivity values. Additionally, a nearest neighbor interpolation algorithm is employed to determine complete geoelectric parameters for the model. Simulation results demonstrate that the ADRN can achieve fast inversion of MT data and yield relatively accurate results. The proposed method has also been applied to invert the COPROD2 dataset and marine MT data from the South Yellow Sea in China. Since field data often contain noise, resulting in discrepancies between real and simulated datasets that may affect the inversion results, noise is added to the input layer of the network to enhance robustness. The results indicate that this method can be effectively applied to field data and shows promising application effectiveness.