Physics-constrained and data-driven synergistic audio-frequency magnetotelluric inversion and its application in gold deposit exploration

Audio-frequency magnetotelluric (AMT) inversion is widely used in deep mineral exploration, but conventional smoothness-constrained inversion can produce overly smooth models, whereas data-driven networks may depend strongly on training data and do not explicitly enforce electromagnetic consistency. This study proposes PD-AMTNet, a physics-constrained and data-driven synergistic AMT inversion framework. PD-AMTNet uses smoothness-constrained inversion results as inversion priors and employs a gated fusion mechanism to integrate these priors with features extracted from two-dimensional (2D) transverse-magnetic (TM)-mode apparent-resistivity and phase data at multiple scales. A physics-driven loss further promotes consistency between the predicted models and their forward responses. Synthetic experiments show that PD-AMTNet yields lower inversion errors and higher structural similarity than smoothness-constrained inversion, U-Net, and U-Net++, with an average RMSE of 0.2052 and an average SSIM of 0.8317 for the tested models. The method also maintains stable performance under the tested noise levels and for model configurations not included in the training dataset. In a field application at the Jinchang gold deposit in southeastern Heilongjiang, China, PD-AMTNet delineates low-resistivity anomalies with sharper boundaries than smoothness-constrained inversion. Two boreholes intersecting known mineralization project near the shallow portions of these anomalies, providing geological support for the interpretation. These results suggest that PD-AMTNet has potential for subsurface resistivity imaging in mineral exploration.