Accurate field-scale estimates of the actual evapotranspiration (ETact) based on readily available input data is indispensable to optimize irrigation in (semi-)arid regions. In this study, at an irrigated pecan orchard in Northwest Mexico we explored the potential for three different machine learning algorithms to improve upon the 30-min ETact estimates provided by the FAO Penman-Monteith method (FAO-PM) when trained and tested on multi-year eddy-covariance measurements. We found that all three machine learning algorithms performed well (R2>0.9), with the Random Forest (RF) giving the best overall performance. RF consistently outperformed FAO-PM on sub-daily scales in the growing season, showing its potential to further improve sub-daily ETact estimates. In addition, RF 1) shows a clear improvement (R2 increase up to 0.168) compared to a typical FAO-PM estimate based on the procedures described in FAO report 56, and 2) mostly performed similar as FAO-PM with a highly accurate locally fitted crop factor on the seasonal scale. The latter highlights its ability to infer the correct seasonal trend through input variables that reflect the underlying physical processes, giving potential for application at other similar fields in the region with limited local tuning. To assess whether incorporating existing (semi-)physical parameterizations could improve the performance, we additionally combined our machine learning models with FAO-PM both with and without locally fitted crop factor. Overall, we found this resulted in a modest improvement in performance in most cases (R2 increase up to 0.01). Notably, incorporating the locally calibrated crop factor and FAO-PM each reduced the errors made by the RF in the second half of the growing season. This implies that it is still useful to incorporate FAO-PM and a crop factor in the machine learning model, especially when the crop factor can be determined with high accuracy.