Abstract
Controlled Source Audio-frequency Magnetotellurics (CSAMT) is widely used in deep resource exploration; however, the systematic distortion of apparent resistivity caused by near-field effects in the low-frequency band has long constrained data interpretation accuracy, and traditional analytical correction methods generally suffer from insufficient accuracy and weak generalization under three-dimensional complex geological conditions. This paper proposes a Frequency-Adaptive Deep Learning Mapping model (FA-DLM) that reformulates CSAMT near-field effect correction as an end-to-end learned mapping task from CSAMT response space to magnetotelluric (MT) response space. The core innovation of FA-DLM lies in the introduction of a frequency-adaptive attention mechanism that establishes cross-band physical coupling dependencies through a log-frequency difference prior bias, combined with a joint optimization strategy of frequency-weighted mean squared error and a phase-consistency auxiliary loss, enabling physically self-consistent joint correction of apparent resistivity and phase. The experimental data were collected independently by the research team in four field survey areas with distinct geological backgrounds (sedimentary basin, carbonate platform, volcanic arc, and thrust belt), comprising 10,810 paired CSAMT-MT synchronous observation records. On the independent test set, FA-DLM achieves the best overall performance (metrics averaged over five independent training runs): apparent resistivity root mean square error (RMSE) of 6.18 [Formula: see text]m, phase RMSE of 3.24[Formula: see text], coefficient of determination [Formula: see text], and near-field reduction rate (NFR) of 85.4%, representing a 7.6-percentage-point increase in near-field reduction rate over the strongest baseline (Transformer without frequency adaptation); the RMSE in the low-frequency high-risk band (0.125-1 Hz) is reduced by 41.2% compared with the non-frequency-adaptive baseline; the inference time is only 18.7 ms per record, satisfying engineering real-time processing requirements. Leave-one-area-out generalization experiments across different geological backgrounds further validate the robustness of the model. This study provides an effective data-driven approach for near-field correction of CSAMT data, with practical value for advancing the accuracy of deep resource exploration.