Despite the success of machine learning (ML) in many disciplines, its application in hydrology, especially in water-scarce regions, faces challenges due to the lack of interpretability and physical consistency. This study addresses these challenges by integrating established empirical hydrological models with ML techniques to predict infiltration rates in water-scarce regions of southern India. Data from 199 observations across 11 sites, including soil characteristics and infiltration measurements, were used to parameterize traditional models like Philip's, Horton's, and Kostiakov's, which were then combined with Artificial Neural Networks (ANN) and the MissForest (MF) algorithm to form hybrid models. The results demonstrate that hybrid models, particularly those based on Philip's model, significantly improve prediction accuracy (R²: 0.76–0.92, RMSE: 0.08–0.2 cm/min, and LCE: 0.11–0.71 with more predictors) across all target sites while retaining interpretability. This approach leverages the strengths of both empirical models and machine learning, addressing the limitations of each. The study highlights that while empirical models are data-driven and may introduce uncertainties, combining them with ML techniques can enhance predictive power and provide a more robust understanding of infiltration dynamics. This is particularly valuable in regions where direct measurement is challenging. The hybrid models facilitate accurate predictions using minimal data from readily accessible locations, offering a practical solution for effective water resource management and soil conservation in semi-arid and data-scarce regions. By blending empirical knowledge with machine learning algorithms, this approach not only improves accuracy but also enhances the physical meaningfulness of hydrological models, providing a balanced and innovative solution to hydrological modeling challenges.