Advancements and Challenges in Assessing and Predicting the Global Carbon Cycle Variations Using Earth System Models

The imperative to comprehend and forecast global carbon cycle variations in response to climate variability and change over recent decades and in the near future underscores its critical role in informing the global stocktaking process. Our study investigates CO₂ fluxes and atmospheric CO₂ growth through ensemble decadal prediction simulations using Earth System Models (ESMs) driven by CO₂ emissions with an interactive carbon cycle. These prediction systems provide valuable insights into the global carbon cycle and, therefore, the variations in atmospheric CO₂. Assimilative ESMs with interactive carbon cycles effectively reconstruct and predict atmospheric CO₂ and carbon sink evolution. The emission-driven prediction systems maintain comparable skills to conventional concentration-driven methods, predicting 2-year accuracy for air-land CO₂ fluxes and atmospheric CO₂ growth, with air-sea CO₂ fluxes exhibiting higher skill for up to 5 years. Our multi-model predictions for the next year, along with assimilation reconstructions, for the first time contribute to the Global Carbon Budget 2023 assessment. We plan regular updates and the involvement of more ESMs in future assessments. Ongoing efforts include implementing seasonal-scale predictions for skill improvement. Furthermore, we assess uncertainty contributions to CO₂ flux and growth predictions, revealing the comparable impacts of internal climate variability and diverse model responses, particularly at a lead time of 1-2 years. Notably, the effect of CO₂ emission forcing rivals internal variability at a 1-year lead time. Large uncertainties in CO₂ responses to initial states of ENSO are observed, stemming from both model responses and internal variability. The challenge lies in addressing the scarcity and uncertainty of data for initialization and obtaining precise external forcings to enhance the reliability of predictions. The further advancements involve not only addressing comprehensive bias correction but also implementing statistical methods to enhance dynamical predictions.