Abstract
Microbiologically influenced corrosion (MIC) accounts for an estimated 20–30% of global corrosion losses, yet reliable quantitative prediction of MIC severity remains unsolved. Electrochemical measurements combined with interpretable machine learning provide a promising framework for addressing this challenge. MIC depends on coupled microbial, biofilm, and interfacial electrochemical processes; charge-transfer resistance (Rct) and corrosion current density (icorr) serve as complementary proxies for corrosion severity, expressed as log₁₀(Rct) from electrochemical impedance spectroscopy (EIS) and log₁₀(icorr) from potentiodynamic polarization. However, physically interpretable predictive models for MIC remain limited. Here, we show that machine learning regressors can predict both log₁₀(Rct) and log₁₀(icorr) in MIC systems involving Pseudomonas and Vibrio across varying environmental conditions. A dataset of 116 EIS and 83 potentiodynamic polarization observations was compiled from ten peer-reviewed sources. Random Forest and Gradient Boosting Machine (GBM) regressors were compared using stratified five-fold crossvalidation, and Shapley additive explanations (SHAP) were used to interpret model behavior in physically meaningful terms. GBM outperformed Random Forest, achieving a higher cross-validated R². SHAP analysis identified double-layer capacitance admittance as the dominant predictor of corrosion severity in the EIS model, consistent with its role as a reporter of biofilm-induced interfacial disorder, while primer-derived organism-level descriptors contributed modest but interpretable signals consistent with known differences in metabolic versatility between the two organisms. This work establishes a reproducible and interpretable baseline for quantitative MIC prediction and, to our knowledge, represents one of the first applications of SHAP analysis to EIS-derived features in MIC while demonstrating the value of integrating organism-level descriptors with electrochemical features for corrosion modeling.