<p>Evaluating subsurface formations through wireline logs is critical for petrophysical analysis in exploration activities yet is often hindered by data quality issues and measurement errors. Addressing these challenges, we propose a novel formation evaluation discrepancy modeling workflow that synergizes machine learning with physics-based models for enhanced well log imputation and interpretation. The workflow is demonstrated on a synthetic dataset that includes standard well logs (e.g., porosity, resistivity, saturation) and facies classifications, generated to represent a geologically realistic environment of alternating marine shales and fluvial or deltaic sandstones. Our methodology leverages physics-based predictions from established principles, such as Gassmann’s equations and Athy’s law, to ensure adherence to known physical laws. We then conduct discrepancy analysis against actual observations to guide feature selection, integrating mutual information and random forest importance ranking. By applying machine learning to predict these discrepancies and combining these predictions on top of our physics-based predictions, a form of machine learning boosting, we refined the initial estimates and significantly improved accuracy. Depth and facies emerge as key predictors in our feature importance analysis, with our discrepancy models combining an initial physics model and random forest model significantly improving prediction accuracy (mean squared error [MSE] of 7.66) over only physics-based predictions (MSE of 18,413). Moreover, Shapley analysis offers additional insights, elucidating the complex influences of model features on predictions and learning the missing physics from the machine learning-based model of discrepancy. Our proposed physics-based discrepancy modeling for well log imputation combines theoretical physics and data-driven predictions to provide more accurate, actionable insights to support reservoir exploration and development decisions.</p>