Water use efficiency (WUE), as a link connecting ecosystem carbon and water cycles, determines carbon sink capacity and water resource sustainability through its variations. However, there is a lack of research on the nonlinear response characteristics and threshold effects of WUE to climate change in arid regions. This study employed interpretable machine learning to systematically analyze WUE-climate relationships in the arid and semiarid regions of northern China, revealing for the first time their nonlinear patterns and critical thresholds. The results show the following: (a) WUE exhibits a marked interdecadal transition characteristic. Before 2010, it showed a decreasing trend (-0.0004 y), while after 2010, it displayed an increasing trend (+0.011 y). This shift is primarily attributed to the weakened influence of drought (Palmer drought severity index) on WUE and the markedly enhanced impact of vapor pressure (VAP). (b) WUE's response to precipitation, mean temperature, and shortwave radiation exhibits nonlinear relationships of increase, increase then decrease, and decrease, respectively, with marked thresholds at similar to 50 mm, similar to 270 K, and similar to 170 W/m(2), respectively. The responses to VAP and Palmer drought severity index are linear, with no marked threshold effects. (c) The spatial distribution of dominant climate factors in these 2 phases shows marked differences in the eastern region, particularly in areas where shortwave radiation is replaced by VAP, which have increased. In summary, the climate response exhibits both dynamic characteristics (varying with phase transitions) and criticality. This implies that ecosystem management in the arid and semiarid regions of northern China needs to shift from "static response" to "dynamic early warning"-as climatic factors approaching thresholds may trigger the risk of abrupt changes.