Small reservoirs in semi-arid regions experience substantial evaporative losses but are rarely monitored at daily scales. A multi-reservoir machine learning (ML) framework was developed to estimate daily open-water evaporation. Empirical models (Penman, Penman-Monteith, Priestley-Taylor, Bowen Ratio Energy Budget) andabenchmark combination method (Daily Lake Evaporation Model-DLEM) were compared against ML models. Predictors combined gridded meteorology (gridMET) with reservoir attributes (surface area, average depth, maximum depth, and fetch). ML models (Random Forest-RF, Decision Tree-DT, K-Nearest Neighbor-KNN, and Support Vector Regression-SVR) were trained on four reservoirs using data from 2018 to 2025. Results from ML models were further validated using both DLEM and TexasETNet observations at Delta Lake. On multi-reservoir tests, RF and SVR perform best (R-2 = 0.67; RMSE similar to 1.5 mm d(- 1); NSE = 0.67). RF aligns most closely with DLEM (R-2 = 0.78; RMSE = 1.22 mm d(-1)), whereas SVR aligns better with TexasETNet (R-2 = 0.33) and shows lowest bias among ML models. Mean annual evaporation from ML and DLEM is 1620-1698 mm yr(-1), which exceeds predictions from TexasETNet (1452 mm yr(-1)) and exhibits narrower interannual variability (+/- 300-350 mm yr(-1) vs. +/- 700 mm yr(-1)). Feature importance and SHAP analyses identified shortwave radiation, temperature (Tmax/Tmin), and ET0 as dominant drivers. The proposed transferable ML framework delivers accurate daily estimates where observations are sparse, while dual validation clarifies differences between process-based and station datasets, supporting reservoir operations and evaporation mitigation planning in under monitored, semi-arid regions.