Abstract
<title>Abstract</title> <p>Electrical resistivity tomography (ERT) pseudosections are not direct images of subsurface resistivity: their responses are distorted by acquisition geometry, sensitivity decay, and the non-uniqueness of the inverse problem. To examine the implications of these effects for karst interpretation, this study develops a forward-modeling-driven framework that explicitly separates physical response analysis from AI-assisted prediction. A ground-truth-complete library of 5,000 synthetic models spanning eight representative karst geoelectrical scenarios was constructed. Apparent-resistivity responses and pseudosections were generated with pyGIMLi using a Wenner-alpha array, and 12 representative models were independently evaluated with RES2DMOD. By comparing forward responses with the known models, we quantified depth compression, lateral displacement, morphological distortion, shallow shielding, and multi-anomaly merging. Boreholes were then simulated using predefined placement strategies, while all borehole observations were extracted exclusively from the ground-truth models. A lightweight U-Net was subsequently used as a proof of concept to predict resistivity and anomaly masks from pseudosections with optional borehole channels. RES2DMOD and pyGIMLi showed close numerical agreement (relative RMS error: 2.03%). The pseudosection responses exhibited substantial geometric and interaction-related distortions. In the synthetic test set, oracle-quality borehole constraints increased anomaly-region intersection over union from 0.114 to 0.506 and reduced the false-positive rate from 0.84 to 0. These findings indicate that forward-response distortion should be quantified before AI-assisted ERT interpretations are assessed. Within the present synthetic design, the benefit of borehole constraints depends strongly on their probability of intersecting the target.</p>