Study region: This study was conducted in a semi-arid Mediterranean region (Rehovot, Israel) within a semi-controlled greenhouse environment characterized by strong spatial microclimatic heterogeneity. Study focus: Accurate estimation of evaporation under heterogeneous microclimates remains challenging, particularly at sub-daily timescales where steady-state formulation often fails. This study aimed to isolate purely physical evaporation dynamics and evaluate whether highfrequency, multi-sensor data combined with machine-learning approaches improve sub-daily evaporation prediction compared to conventional FAO-56-Penman-Monteith. Open-water evaporation was continuously measured using 62 high-throughput gravimetric pans, paired with 12 spatially distributed meteorological stations sampled every 3 min over one year (similar to 2.4 million records). New hydrological insights for the region: Spatial heterogeneity reached a temperature gradient of 8 degrees C and relative humidity differences of 16%. Strong hysteresis was observed, whereby identical atmospheric drivers produced different evaporation rates depending on the time of day. Temporal encoding (hour of day and day of year) and light intensity dominated predictive importance, while vapor pressure deficit contributed minimally. XGBoost achieved the highest accuracy (R-2 = 0.935) and remained robust with 10% of the dataset (R-2 = 0.935). Tree-based models consistently outperformed Neural-Network, which showed higher sensitivity to data reduction. FAO-56 Penman-Monteith showed substantially lower performance (R-2 = 0.637).Together, these results show that hysteresis and temporal structure are critical for accurate high-frequency evaporation prediction under complex, non-steady microclimates, motivating machine-learning frameworks that can learn lagged, non-linear responses.