Deep Neural Networks for Atmospheric Transport Modeling of CO2

Large deep neural network emulators are poised to revolutionize numerical weather prediction (NWP). Recent models like GraphCast or NeuralGCM can now compete and sometimes outperform traditional NWP systems, all at much lower computational cost. However, the potential applicability of these emulators to other dense prediction tasks, such as modeling 3D atmospheric composition, remains largely unexplored. In this study, we introduce a novel approach to atmospheric transport modeling, focusing specifically on CO2 and other inert trace gases. Unlike existing Eulerian transport modeling methods that rely on computationally expensive numerical solvers applied to the continuity equation, our approach leverages neural networks trained in latent space. By doing so, we aim to achieve high-fidelity transport modeling while significantly mitigating computational costs. To facilitate the training and evaluation of our neural network-based approach, we have developed the CarbonBench dataset, which incorporates atmospheric CO2 inversion data, corresponding meteorological reanalysis data, and station observations. Through qualitative and quantitative experiments, we have identified challenges in training different neural network architectures, with mesh-based methods (graph neural networks and u-nets) proving to be particularly difficult to train, in contrast to spectral approaches (spherical fourier neural operators). However, our model demonstrates promising results in reducing input-output (IO) load by over 4x, thereby enhancing the efficiency of forward runs, especially in inverse modeling frameworks. This advancement opens up new possibilities for integrating ecosystem surface fluxes into carbon cycle inversions, which has not been previously explored. By presenting the first global Eulerian atmospheric transport emulator, our work represents a significant step towards advancing carbon cycle inverse modeling using deep learning techniques. This lays the groundwork for future research aimed at optimizing flux models with a greater number of parameters, ultimately enabling more comprehensive and accurate carbon cycle predictions.