Quality Control of Magnetotelluric Impedance Inversion via Self‐Supervised Learning Based on Data Continuity of Nyquist Diagrams

ABSTRACT Magnetotelluric (MT) data are often contaminated by electromagnetic (EM) noise, making quality control (QC) essential for reliable inversion. Conventional manual time‐domain QC processes are labour‐intensive, and some noise inevitably remains after processing, necessitating complementary processing in the frequency domain to address the residual noise. We propose a novel self‐supervised learning framework that serves as an additional post‐processing QC step performed in the frequency domain after QC involving direct data editing. This method mitigates the influence of noisy measurements during inversion by adaptively assigning lower weights to frequency bands that deviate from the expected geometrical patterns on the Nyquist diagram. To directly handle the sparse, point‐cloud representation of frequency‐domain MT data, we employ the Point Transformer architecture, which is specialized for point‐cloud data. During training, the network is guided by the geometrical patterns of MT data on the Nyquist diagram. In addition, we introduce two supplementary strategies–‘ loss’ to regularize training and ‘logarithmic scaling’ to linearize the frequency–impedance relationship. Numerical experiments show that the proposed QC approach effectively identifies outliers and improves inversion performance. The supplementary strategies further enhance outlier detection and training stability. A key advantage of our framework is its flexibility, as the proposed QC method can be applied independently to each MT station, or even more granularly, to each frequency band.