<p>Accurate estimation of water balance components is fundamental for effective water resource and agricultural management, particularly in in water-scarce regions. Among these components, actual evapotranspiration (Eₐ) remains the most uncertain yet critical variable. This study evaluates the performance of widely used Budyko-like conceptual curves alongside two machine learning (ML) approaches, Random Forest and XGBoost, across 598 sub-basins in Iran, representing diverse physiographic and climatic conditions in both “Plain” and “Highland” areas. By integrating sensitivity analysis, calibration, and uncertainty assessment, we quantify the influence of key hydro-climatic and physiographic factors, including precipitation, slope, vegetation condition (NDVI), and geographic location. Model performance was evaluated using R², NSE, RMSE, KGE, and uncertainty metrics including average relative interval length (ARIL) and prediction level (P_level). ML models substantially outperformed conceptual approaches, with XGBoost achieving the highest validation accuracy (NSE = 0.61, RMSE = 0.091), followed by RF (NSE = 0.58, RMSE = 0.096). Sensitivity and explainability analyses reveal that slope and dryness index are the dominant controls on model residuals and predictive skill. Uncertainty analysis further reveals that while Budyko formulations such as Zhou’s curve exhibit wide prediction intervals, ML models more effectively capture observed variability, albeit with broader uncertainty bounds. The findings highlight the complementary strengths of conceptual and data-driven frameworks and emphasize the importance of incorporating topographic and climatic gradients for robust evapotranspiration modeling in data-scarce, hydro-climatically complex regions. </p>