Potential evapotranspiration (PET) is a key variable in drought occurrence and modeling. The no-water-limited Bowen ratio (beta NWL) is widely used to construct energy balance-based PET models by assuming that beta NWL does not vary with climate and vegetation conditions. However, we found that beta NWL varies significantly with climate as well as vegetation conditions based on global-wide observational flux data. Therefore, this study aims to investigate the dominant influence factors of beta NWL and to simulate the nonlinear relationship between beta NWL with environmental factors by leveraging flux observation globally and machine learning models. We then applied the nonlinear beta NWL to develop a PET model (PET beta NWL-RF ) and evaluated its performance under various conditions, comparing it against commonly used PET models. Our results showed that the gross primary productivity (GPP) had the most significant effect on beta NWL, with a relative importance of 31%. The PET beta NWL-RF model significantly improved the accuracy of daily PET estimation (R2 >= 0.93, TSS >= 0.96, RMSE <= 0.48 mm/day, -0.04 mm/day <= MB <= 0.06 mm/day) against observation. Moreover, we also found that the PET beta NWL-RF model can effectively reduce the uncertainty (overestimation or underestimation) of PET estimation by commonly used PET models especially under drought conditions and hence significantly enhance the reliability of drought monitoring. This study reveals the influence of nonlinear relationships of surface energy partitioning on PET, which would be insightful for PET estimation as well as drought monitoring.