Abstract The Three-North Shelter Forest Program (TNSFP) of China, one of the largest forestry ecological projects, serves as a nature-based solution to addressing ecological, social, and economic challenges and issues. This program, mainly situated in arid and semi-arid regions, critically focuses on improving water use efficiency (WUE) - maximizing carbon sequestration per unit of water consumed - as a key strategy for optimizing water resource utilization and ensuring the long-term success of vegetation restoration efforts. However, the regulatory mechanisms of forest WUE in this region remain unclear. Here, we adopted an interpretative machine-learning method to investigate the influence of main environmental elements, topographical conditions, and stand characteristics on forest WUE in the TNSFP region from 2001 to 2022 based on remote sensing products, ground monitoring data, and forest inventory data. Our study identified soil moisture (SM) as the primary factor influencing forest WUE across the TNSFP region, with higher SM levels generally leading to improved WUE in forests. However, stand characteristics strongly mediated their relationship. Specifically, forest WUE initially increases against forest density before peaking at about 1000 trees hm–2 for needle-leaved forests and 800 trees hm–2 for broad-leaved forests, respectively, then gradually declining due to water competition. When SM is relatively adequate, moderate thinning could significantly enhance forest WUE. Furthermore, implementing management strategies to improve WUE is crucial as needle-leaved forests mature. This study emphasizes the significant impacts of stand characteristics on forest WUE in the TNSFP region, offering essential insights for optimizing water resource management in managed forests across arid and semi-arid regions.