<p>Water permeability governs the ingress of aggressive agents, such as chlorides and sulfates, critically affecting concrete durability and service life. While several machine learning (ML) studies have explored the short-term laboratory datasets, few have examined long-term field behavior under diverse environmental exposure conditions. This study leverages an unprecedented 19-year field database comprising 1924 data points across tidal, aboveground, and belowground exposure zones in a marine environment to predict and interpret the evolution of water permeability in concrete with varying composition. The effect of key parameters, including supplementary cementitious materials (SCMs), fibers, corrosion inhibitors, curing methods, and water type, on water permeability in concrete were evaluated. Three ML models, namely support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost), were trained and compared. XGBoost exhibited the best generalization (<i>R</i>² = 0.87), capturing non-monotonic permeability trends consistent with cyclic wetting and drying. Model interpretability was enhanced through SHAP-based feature importance and correlation analysis, revealing the dominant influence of SCM type, w/cm ratio, and exposure conditions. This work bridges experimental field evidence with interpretable ML prediction, offering a robust, data-driven framework for durability design and potential integration with reliability or service life models.</p>