CONTEXT: Global water scarcity challenges sustainable development, particularly in arid and semi-arid regions where agriculture consumes the majority of available freshwater. Accurately evaluating water-saving potential a regional scale is essential for informed water resource allocation. However, upscaling from the field to the regional level remains a significant challenge due to the spatial complexity of environmental factors. OBJECTIVE: This study developed a framework to estimate the regional agricultural water-saving potential by integrating a process-based crop growth model with machine learning. Focusing on the hyper-arid Turpan-Hami Basin, we quantified the water-saving potential of major field crops and analyze the effectiveness of various irrigation optimization scenarios. METHODS: We parameterized the AquaCrop model using 140 field sites to simulate yield and water-use responses for six major crops (grape, cotton, cantaloupe, sunflower, maize, and wheat) under varying Total Available Water (TAW) thresholds. A Differential Evolution (DE) multi-objective optimization algorithm was applied to optimize irrigation thresholds to minimize water consumption while maintaining yield. Subsequently, we used a CatBoost machine learning model and environmental covariates from Google Earth Engine to upscale the field-scale water-saving potential to the regional level. RESULTS AND CONCLUSIONS: Local validation against canopy cover (CC) and yield observations demonstrated robust model performance across the main crops (wheat, maize, sunflower, cotton, grapes, and cantaloupe) in this study, with in-season RMSE values ranging ranging from 5.8% to 23.8% for CC and from 0.06 t ha(-1) to 0.52 t ha(-1) for yield. Based on these calibrated models, AquaCrop simulations indicated substantial variations in the water-saving potential: surface-irrigated grapes exhibited the highest average potential (2967.5 m(3) ha(-1)), whereas micro-irrigated cotton showed the lowest average potential (15.9 m(3) ha(-1)). The CatBoost model achieved high predictive accuracy (R-2 = 0.89, MAE = 222.60 m(3) ha(-1)) for regional estimation. Feature importance analysis identified crop type, irrigation method, temperature, elevation, and dew point temperature as the primary drivers of water-saving potential. We concluded that optimizing irrigation thresholds under current practices could save 1.075 x 10(8) m(3) across the basin. This figure could increase to 1.827 x 10(8) m(3) if micro-irrigation were fully adopted. SIGNIFICANCE: This study highlights that combining crop modeling with machine learning effectively bridges the gap between field-scale experiments and regional-scale water assessments. Our results suggest that upgrading irrigation systems and optimizing irrigation scheduling are more feasible water-saving strategies in arid regions than altering established cropping patterns. These findings provide data-driven guidance for sustainable water management in water-scarce basins.