Physics-Guided Unsupervised Deep Learning Layered Earth Inversion of Frequency-Domain CSAMT Data for Simultaneous Resistivity and Induced Polarization Parameter Recovery

Induced polarization (IP) affects controlled-source audio-frequency magnetotelluric (CSAMT) data over sulfide-bearing rocks. Recovering resistivity together with the Cole–Cole IP parameters is difficult because the parameters trade off against each other and conventional inversion must be restarted at every sounding. We present PGUD-IP (physics-guided unsupervised deep learning for induced polarization), a neural network that maps the impedance of one sounding to a layered Cole–Cole model. It is trained without true models: each predicted model passes through a one-dimensional forward calculation, and training minimizes the misfit to the input impedance. On 200 noisy three-layer synthetic models, PGUD-IP and conventional least-squares inversion with the same forward calculation reached relative impedance errors of 0.173 and 0.194. PGUD-IP recovered resistivity, relaxation time, and thickness better, and the conventional inversion recovered chargeability and the frequency exponent better. A network trained only on synthetic ten-layer models was then applied to 37 soundings of line L01 in the Quaternary-covered periphery of the Geyubao gold deposit, Hebei Province, China. Without fitting the field data, it matched the measured responses as closely as conventional inversion (relative error 0.437 versus 0.439) in 0.02 s instead of 47 s. An inversion without IP fitted almost as well (0.443), but a borehole at about 640 m intersected gold-bearing pyritized cataclastic rocks at 200–240 m, within the high-resistivity, high-chargeability anomaly recovered by PGUD-IP. PGUD-IP is suited to fast one-dimensional resistivity imaging and to delineating concealed IP targets.