<p>Magnetotellurics (MT) investigates the subsurface resistivity from naturally occurring electromagnetic field fluctuations. The linearity and noise in MT data cause traditional inversion methods to fail, resulting in poor resistivity reconstruction. While convolutional and recurrent neural networks have been applied to invert MT, they cannot be used to include complex, non-local relationships over depth and frequency. In this research, we propose a hybrid model that combines graph attention networks (GATs) and multihead attention mechanisms. The GAT module serves as both denoiser and encoder, structuring the input frequency–phase data into a compressed representation, and the transformer-based attention selectively emphasizes the most informative spectral elements. This allows the model to predict resistivity distributions more accurately with depth. Experiments on both synthetic and field datasets demonstrate that the proposed model outperforms the Convolutional Neural Network (CNN) Mean Squared Error (MSE) (0.0030), with correlation coefficient R = 0.78, and Vanilla Neural Network (VNN) MSE (0.0173), with correlation coefficient R = 0.71 baselines with a preeminence of MSE of 0.0018 and a correlation coefficient of 0.85. The integration of graph attention with frequency-aware multihead attention allows the current model to capture the underlying physics of MT data in a far more effective way and provides resistivity profiles that are accurate and less noisy.</p> Graphical Abstract <p>This graphical abstract summarizes a study on the one-dimensional (1D) inversion of magnetotelluric (MT) data using a novel deep learning model that combines Graph Attention Networks (GATs) with multi-head attention mechanisms. The GATs denoise and encode the MT data by structuring it as a graph, which captures dependencies between frequencies. The multihead attention component then dynamically focuses on the most informative frequency elements to predict subsurface resistivity distributions more accurately with depth. Evaluated on both synthetic and real-world datasets from the United Arab Emirates, the model demonstrated superior robustness to noise and better generalization than current deep learning models. This approach offers a more scalable, interpretable, and physically informed solution for MT inversion.</p>