Human activities release vast amounts of carbon dioxide (CO₂) and other greenhouse gases into the atmosphere, leading to significant increase in global surface temperature (climate warming) and associated alterations in global precipitation patterns. The terrestrial ecosystem plays a crucial role in regulating Earth's climate by taking up about 1/3 of the anthropogenic CO₂ emission annually. On the other hand, climate change exerts profound impacts on ecosystem carbon cycling, altering the rates of carbon uptake and release by terrestrial ecosystems. For instance, warmer temperature accelerates carbon decomposition processes in soils, potentially leading to increased CO₂ emissions. Additionally, changes in precipitation patterns may affect vegetation growth and distribution, influencing the capacity of plants to sequester carbon through photosynthesis. In the meantime, human activities such as industrial processes, agricultural practices, and fossil fuel combustion have caused global nitrogen (N) deposition, a significant environmental issue with far-reaching implications for climate change through influencing carbon cycling.  Thus, understanding the intricate interplay between climate change and carbon cycling is crucial for developing effective strategies to mitigate the impacts of climate change on Earth's sustainability.

Although terrestrial carbon cycling provides critical feedback to climate change, it remains the most uncertain element in Earth System Models (ESMs). Comparing model outputs with large and diverse datasets enables the correction of biases and the reduction of uncertainty in model predictions. Our overarching goal of this proposal is to use experimental benchmarks to understand, quantify, and reduce model uncertainty and correct model biases in the land carbon cycle.

To achieve this goal, we propose three objectives: 

Obj.1. Derive new benchmarks and metrics from global change experiments and integrate them into the International Land Model Benchmarking (ILAMB) package for ESM evaluation and development. We will assemble data from the thousands of global change experiments in which global change factors (i.e., atmospheric CO₂ concentration, air/soil temperature, precipitation regime, and nitrogen deposition) are manipulated, and carbon variables are measured in the field. The benchmarks include mean responses of the carbon cycle to GCFs in major biomes and the temporal trend of these mean responses in those decade-long experiments. The emergent functional relationships between the responses and environmental factors, including climate, vegetation, and soil conditions, will also serve as benchmarks. These functional relationships will include individual and collective spatial relationships between carbon variables and environmental factors. Individual relationships will be determined using advanced regression analysis, and the collective relationships will be determined through machine learning. Multiple machine learning algorithms such as Random Forest, XGBoost, and Neural Network will be deployed to determine the functional relationships and assess the relative importance among climate, vegetation, and soil. Both the biome-dependent mean responses and functional relationships will be integrated into ILAMB for model benchmarking. 

Obj.2. Design a new model intercomparison project (MIP) for comparisons with the experimental benchmarks derived in Obj.1. In parallel to field experiments, we will design model perturbation experiments in which environmental forcings to land models are manipulated the same way as those in global change experiments. The mean responses to GCFs and functional relationships derived in Obj.1 will be compared with the outputs of the model perturbation experiments. The ESMs deployed for the model perturbation experiments include ELM (the highest priority) in E3SM, CLM in CESM, CABLE in ACCESS, and ORCHIDEE in IPSL. 

Obj.3. Develop an online hybrid model, in which machine learning is coupled with ELM for an end-to-end parameterization to correct the biases and reduce the uncertainties revealed in Obj.2. Deep neural networks (NNs) will be deployed as surrogates for the processes with large biases or uncertainties identified in Obj. 2. These NNs will couple with ELM to form a hybrid model which can be trained end-to-end with experimental benchmarks. This hybrid model will serve as a component for ELM online operation to correct the biases and reduce uncertainties in E3SM predictions. In the meantime, key model parameters will be retrieved via these NNs for further analyses, such as comparison with default values. This project is expected to further develop the ILAMB software package and improve the ESM model performance in terrestrial carbon cycling.