Tracing and Predicting the Pulse of the Terrestrial Biosphere: Plant Physiology, Earth System Models, and AI

Earth System Models (ESMs) remain our primary tools for predicting future climate, yet they face significant challenges and uncertainties in reliably resolving the terrestrial carbon-water dynamics under a changing climate. In this presentation, I share three integrated efforts within the NCAR Community Earth System Model (CESM) framework that bridge plant physiology and artificial intelligence (AI) to advance our understanding of the global carbon sink and the carbon-water cycle interactions. First, we address a long-overlooked but fundamental plant physiological process: the mesophyll diffusion of CO2 inside leaves, which critically determines photosynthesis, transpiration, and water use efficiency (WUE). We developed the first global mesophyll conductance model, and revealed that without explicit representation of this internal leaf physiology process, the global CO2 fertilization effect can be significantly underestimated by ESMs. We then extended this mechanistic foundation to novel photosynthetic tracers, OCS and 13C isotopes, which only accounts for stomatal conductance in current ESMs. We found that the mesophyll effect largely accounts for the global secular increase in 13C/12C isotopic discrimination of photosynthesis under elevated CO2, and contributes to a proportional increase in intrinsic WUE (iWUE) to atmospheric CO2 rises and therefore steady increase of global ecosystem WUE (eWUE). This increase occurs because the CO2 fertilization effect continuously outweighs the gradually strengthening VPD effect. The eWUE gain is projected to continue through 2100 under moderate to high emission scenarios. Our findings suggest that terrestrial ecosystems will continue to benefit from the water-saving effect under climate change. In the third study, we developed a novel deep learning (DL) framework based on the Autoformer architecture to constrain ESM uncertainties in future projections of terrestrial carbon dynamics using CMIP6 model ensembles. This DL framework goes beyond improving computational efficiency of ESMs as is most often pursued so far, but aims to learn the structure dependence of the model-data deviation: the aggregated outcome of existing unresolved processes and parameters. We then applied the correction of model–data deviation learned from the past to the future. Our results reveal a likely overestimation of future carbon sequestration capability by the current ESMs. This finding suggest that a sustained future increase in the terrestrial carbon sink is not guaranteed. Meanwhile, DL has markedly tightened the uncertainty range of existing ESM projections, and therefore can enhance the fidelity of future projections of carbon cycling and guide global climate actions.