A physics-guided machine learning approach to screening spectral induced polarization measurements

Spectral induced polarization (SIP) measurements are increasingly used in time-lapse monitoring studies and require consistent quality control before interpretation. SIP spectra are information-rich but susceptible to acquisition artifacts, noise, and nonphysical behavior, making reliable screening essential before spectra are used to infer material or process changes. Existing approaches rely largely on manual expert review or automated classifiers that do not explicitly evaluate physical plausibility. Here, we develop a physics-guided machine-learning workflow for raw SIP quality control that combines CatBoost classification with spectral-shape checks and Kramers-Kronig consistency diagnostics. The workflow was developed from 599 spectra, including 107 labeled spectra used for grouped validation, and assigns spectra to Good, Bad, or Review categories so uncertain cases are retained for inspection rather than forced into binary classifications. In grouped out-of-fold evaluation, the CatBoost-only workflow achieved an auto-accept precision of 0.80 and a Bad escape rate of 0.02, whereas the physics-guided workflow achieved 1.00 auto-accept precision and reduced the Bad escape rate to 0.00 by routing uncertain or physically questionable spectra to Review or Bad. Although conservative, the physics-guided workflow improved resolved accuracy from 0.47 to 0.59 among spectra assigned final Good or Bad labels. These results show that the main value of the workflow is not higher classification accuracy alone, but a defensible quality-control framework that combines statistical pattern recognition with physical screening. This framework supports scalable SIP data validation and future time-lapse monitoring of large temporal datasets in which ongoing data streams can be screened consistently before meaningful temporal changes are interpreted.