The current study evaluated the accuracy of four machine learning (ML) techniques and thirteen experimental methods calibrated to estimate potential evapotranspiration (ET0) in arid and semi-arid regions. Various scenarios utilizing meteorological data were examined, and FAO56-PM was used as a benchmark. The results revealed that the ML models outperformed the experimental methods at both daily and monthly scales. Among the ML models, the 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 the ANNs, SVM, RF, and GAM exhibited greater accuracy on a daily scale. Notably, the ANNs and SVM achieved high accuracy even with a limited number of variables. Conversely, the accuracy of the RF improved with an increased number of variables. Comparing ML models to experimental models with equivalent input revealed that ANN with inputs similar to Valiantras-1 performed better on a monthly scale, while SVM with inputs akin to Valiantras-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.