ABSTRACT Ecosystem exchanges of carbon, water, and energy are central to Earth system functioning, yet their sensitivities to environmental variability remain poorly constrained across biomes and climates. Here, we analyzed ≥ 5 years of eddy covariance data from 87 AmeriFlux sites (964 site‐years) spanning six vegetation types and a broad range of climatic conditions to examine the controls and multi‐year trends of gross primary productivity (GPP), evapotranspiration (ET), water‐use efficiency (WUE), and the Bowen ratio. We trained boosted regression tree ensembles with environmental (air temperature, vapor pressure deficit, soil water content, atmospheric CO 2 , radiation, wind speed) and temporal (month, year) variables and used interpretable machine learning—SHapley Additive exPlanations (SHAP) and Accumulated Local Effects (ALE)—to quantify driver importance and nonlinear responses. Across biomes and along climatic gradients, environmental controls reorganized predictably: radiation dominated under warm and humid conditions, whereas soil moisture exerted a stronger influence in drier or warmer systems. GPP and ET were shaped by similar dominant drivers, with radiation and temperature generally strongest but soil moisture exerting comparable influence in some biomes. WUE was consistently constrained by vapor pressure deficit, indicating stomatal regulation under rising atmospheric dryness. In contrast, controls on the Bowen ratio diverged more across biomes, indicating heterogeneity in how ecosystems partition energy between latent and sensible heat. Multi‐year trend analysis revealed negative associations of GPP and ET with rising atmospheric demand, and positive associations of the Bowen ratio with drying, indicating reduced evaporative cooling and stronger land–atmosphere coupling. Together, these findings show that while carbon and water fluxes remain tightly coupled across timescales, their balances reorganize predictably along bioclimatic gradients. This framework underscores the value of long‐term flux networks and interpretable machine learning for benchmarking Earth system models and constraining projections of terrestrial carbon–water–energy dynamics under climate change.
Yi et al. (Fri,) studied this question.