The current study evaluated the accuracy of four machine learning (ML) models and thirteen experimental methods calibrated to estimate reference evapotranspiration (ET0) in arid and semi-arid regions. Various scenarios were examined utilizing meteorological data and FAO56-PM as a benchmark. According to the results, the ML models outperformed the experimental methods on both daily and monthly scales. Among the ML models, artificial neural networks (ANNs), generalized additive model (GAM), random forest (RF), and support vector machine (SVM), respectively, demonstrated higher accuracy on a monthly scale, while ANNs, SVM, RF, and GAM exhibited greater accuracy on a daily scale. Notably, ANNs and SVM achieved high accuracy even with a limited number of variables. Conversely, RF showed improved accuracy with an increased number of variables. Comparing the ML and experimental models with equivalent inputs revealed that ANN with inputs similar to Valiantzas-1 performed better on a monthly scale, while SVM with inputs akin to Valiantzas-3 showed superior performance on a daily scale. Our findings suggest that average temperature, wind speed, and sunshine hours contribute significantly to the accuracy of ML models. Consequently, these ML models can serve as an alternative to the FAO56-PM method for estimating ET0.