Graph Attention Networks with Multihead Attention for Improved Resistivity Model Estimation

Magnetotellurics (MT) investigates the subsurface resistivity from naturally oc curring electromagnetic field fluctuations. The linearity and noise of MT data cause traditional inversion methods to fail due to the nonlinearity and noise inherent in MT data, resulting in poor resistivity reconstruction. While neural networks such as convolutional and recurrent networks have been applied to in vert MT, they cannot be used to include complex, non-local relationships over depth and frequency. In this research, we propose a hybrid model that com bines graph attention networks (GATs) and multi-head attention mechanisms. The GAT module serves the double purpose of denoiser and encoder, structur ing the input frequency-phase data into a compressed representation, and the transformer-based attention selectively emphasizes the most informative spec tral elements. This allows the model to predict resistivity distributions more accurately with depth. Experiments on both synthetic and actual datasets show enhanced robustness to noise and enhanced generalization over current deep learning models, providing a more interpretable and scalable solution to MT inversion.