Using machine learning models (MLMs) to develop high-accuracy evapotranspiration (ET) products is important for investigating the terrestrial hydrological changes in arid and semi-arid regions in the context global warming.Based on the 12 flux stations in Northwest China and multi-source observation datasets, we present a 5-km gridded ET product based on 4 MLMs including the random forest, the extreme gradient boosting, the support vector regression, and the artificial neural network, and analyze the long-term ET trend over Northwest China.The cross-validation results show that all the four models can simulate the daily ET reasonably well, with the root-mean-square error (RMSE) smaller than 0.57 mm·d^-1and the R² up to 0.73~0.88.Moreover, the Sharply additive explanations (SHAP) method reveals that all the models treat the net radiation, vegetation indexes and soil moisture as the most important predictors and capture the limitation effect of soil water on ET reasonably well, indicating a good physical interpretability of the 4 MLMs.No model always has superiority, and the ensemble mean of the 4 models shows a 7%-20% and 45%-70% smaller RMSE than the individual member and other ET products.The ensemble ET shows an increasing trend over the Northwest China during 2001-2018, with a mean increase of 19 mm/(10 a).In addition, the rate of growth of ET is greater than the rate of increase of precipitation in the Hetao region and the middle and northeastern parts of Inner Mongolia, suggesting an intensified drying trend in these regions.