Atmospheric Transport Modeling of CO 2 With Neural Networks

• Published in: Journal of advances in modeling earth systems Vol. 17; no. 2
• Main Authors: Benson, Vitus, Bastos, Ana, Reimers, Christian, Winkler, Alexander J., Yang, Fanny, Reichstein, Markus
• Format: Journal Article
• Language: English
• Published: 01.02.2025
• ISSN: 1942-2466; 1942-2466
• DOI: 10.1029/2024MS004655

Accurately describing the distribution of in the atmosphere with atmospheric tracer transport models is essential for greenhouse gas monitoring and verification support systems to aid implementation of international climate agreements. Large deep neural networks are poised to revolutionize weather prediction, which requires 3D modeling of the atmosphere. While similar in this regard, atmospheric transport modeling is subject to new challenges. Both, stable predictions for longer time horizons and mass conservation throughout need to be achieved, while IO plays a larger role compared to computational costs. In this study we explore four different deep neural networks (UNet, GraphCast, Spherical Fourier Neural Operator and SwinTransformer) which have proven as state‐of‐the‐art in weather prediction to assess their usefulness for atmospheric tracer transport modeling. For this, we assemble the CarbonBench data set, a systematic benchmark tailored for machine learning emulators of Eulerian atmospheric transport. Through architectural adjustments, we decouple the performance of our emulators from the distribution shift caused by a steady rise in atmospheric . More specifically, we center input fields to zero mean and then use an explicit flux scheme and a mass fixer to assure mass balance. This design enables stable and mass conserving transport for over 6 months with all four neural network architectures. In our study, the SwinTransformer displays particularly strong emulation skill: 90‐day and physically plausible multi‐year forward runs. This work paves the way toward high resolution forward and inverse modeling of inert trace gases with neural networks. Changes in the concentration can be measured in our atmosphere. To connect these to emissions, and activity from biosphere and ocean ecosystems, traditionally an atmospheric transport model is used that tracks the flow of with the winds. Now, with progress in artificial intelligence (AI), it can be questioned, if these atmospheric transport models can be replaced with an AI model. In this work we introduce CarbonBench, a benchmark data set to train and compare different AI models. Moreover, we design a state‐of‐the‐art AI model to predict how distributes in the atmosphere. All our data and code are open‐source, with the aim to enable further research toward leveraging AI for monitoring greenhouse gases and supporting climate agreements. CarbonBench: a systematic benchmark for machine learning emulators of atmospheric tracer transport Adapted SwinTransformer deep neural network to achieve stable and mass‐conserving transport of by including physical constraints UNet, GraphCast, and Spherical Fourier Neural Operator baselines with the same customization are also strong models, for shorter lead times (up to 90 days)