Understanding the exchange of CO2 and other greenhouse gases (GHGs) between land, atmosphere, and ocean is crucial for mitigating climate change and supporting climate agreements. However, significant uncertainties remain due to challenges in integrating experimental, observational, and theoretical research across scales. Data-driven machine learning (ML) approaches have become popular for studying different components of the carbon cycle, but current artificial intelligence (AI) systems rarely provide a comprehensive view of the entire Earth system. This session aims to connect diverse research communities to discuss AI-driven research on the carbon cycle.
We encourage submissions on all aspects of the carbon cycle, including the atmosphere, biosphere, ocean, and human impacts. ML can enhance top-down and bottom-up approaches for quantifying land and ocean fluxes, constraining carbon budgets and carbon stocks, and mapping CO2 and CH4 through atmospheric tracer transport. This is crucial for tasks such as partitioning land fluxes into photosynthesis and respiration, estimating carbon stocks in soils and biomass, etc. This session particularly targets works that integrate diverse data sources that are not traditionally combined, such as remote sensing data with eddy covariance flux measurements.
Recent advances in numerical weather prediction have shown the effectiveness of deep neural networks, particularly transformers. Emerging "Foundation Models" integrate diverse data streams to address multiple tasks within a single AI system. Classical algorithms like random forests, Gaussian processes, and gradient-boosting trees remain useful due to their efficiency and ease of use. Uncertainty quantification and explainable AI enhance ML methods by revealing key variables.
In summary, this session welcomes all submissions that leverage machine learning for carbon cycle research, including but not limited to:
- Machine learning for integrating remote sensing data with in-situ observations.
- Hybrid modeling for enhanced inference of parameters and responses of key processes in the carbon cycle.
- Mapping of stocks, fluxes, and other carbon-related quantities.
- Emulators to speed up conventional models, such as atmospheric tracer transport models for GHGs.
- ML methodologies for improving uncertainty quantification, prediction accuracy, explainability, or sample efficiency.
- Model-data integration to better understand CO2 and CH4 fluxes.
We encourage submissions on all aspects of the carbon cycle, including the atmosphere, biosphere, ocean, and human impacts. ML can enhance top-down and bottom-up approaches for quantifying land and ocean fluxes, constraining carbon budgets and carbon stocks, and mapping CO2 and CH4 through atmospheric tracer transport. This is crucial for tasks such as partitioning land fluxes into photosynthesis and respiration, estimating carbon stocks in soils and biomass, etc. This session particularly targets works that integrate diverse data sources that are not traditionally combined, such as remote sensing data with eddy covariance flux measurements.
Recent advances in numerical weather prediction have shown the effectiveness of deep neural networks, particularly transformers. Emerging "Foundation Models" integrate diverse data streams to address multiple tasks within a single AI system. Classical algorithms like random forests, Gaussian processes, and gradient-boosting trees remain useful due to their efficiency and ease of use. Uncertainty quantification and explainable AI enhance ML methods by revealing key variables.
In summary, this session welcomes all submissions that leverage machine learning for carbon cycle research, including but not limited to:
- Machine learning for integrating remote sensing data with in-situ observations.
- Hybrid modeling for enhanced inference of parameters and responses of key processes in the carbon cycle.
- Mapping of stocks, fluxes, and other carbon-related quantities.
- Emulators to speed up conventional models, such as atmospheric tracer transport models for GHGs.
- ML methodologies for improving uncertainty quantification, prediction accuracy, explainability, or sample efficiency.
- Model-data integration to better understand CO2 and CH4 fluxes.