AQcGAN: Deep Generative Emulation of NASA GEOS Atmospheric Composition Across Space, Time, and the Full Atmospheric Column

Presenter: Jennifer Sleeman
Organization: Johns Hopkins Applied Physics Laboratory
Co-Authors: Jennifer Sleeman (Co-PI), Viral Shah (PI), Patricia Castellanos (Co-I), Christopher James Ribaudo, Adeline Hillier, Caroline Tang, Collin Kofroth, Alex Chen, Pam Wales, Pratik Shukla, Jules Kouatchou, K. Emma Knowland, Christoph Keller

Abstract

The NASA GEOS-FP (Forward Processing) and GEOS-CF (Composition Forecast) systems provide global analyses and forecasts of atmospheric composition for air quality research, chemical data assimilation, and operational NASA applications. However, simulating atmospheric chemistry, aerosol processes, and tracer transport is computationally expensive, limiting forecast length, ensemble size, and uncertainty quantification. These challenges will become increasingly important as future aerosol data assimilation systems transition to ensemble-based methodologies.

We developed the Air Quality conditional Generative Adversarial Network (AQcGAN), a deep learning framework that emulates atmospheric composition in GEOS-FP and GEOS-CF. AQcGAN produces rapid 10-day ensemble forecasts of reactive gases and aerosols while requiring only a small fraction of the computational cost of the underlying numerical models. Unlike many deep learning forecasting approaches, AQcGAN employs a hybrid sequence-to-sequence autoregressive framework that incorporates future meteorological and emissions information to improve forecast skill.

AQcGAN has recently been extended to emulate the full 72-level atmospheric column. The framework predicts major trace gases (CO, NO, NO₂, and O₃) and aerosol species including dust, sea salt, sulfate, nitrate, black carbon, and organic carbon. Architectural advances include double convolution networks, a spherical loss function that accounts for latitude-longitude grid geometry, and an autoencoder-based AQcGAN (AE-AQcGAN) that exploits correlations across vertical levels.

AE-AQcGAN reduced forecast RMSE for gaseous species relative to the original single-level model, with the largest improvements occurring in the middle and upper atmosphere. For aerosols, forecasts of 19 mass mixing ratio species consistently outperformed persistence for most species at 10-day lead times. These results demonstrate that deep generative models can accurately emulate the vertically coupled evolution of atmospheric composition while enabling rapid forecast generation and large ensembles for uncertainty quantification that would otherwise be computationally prohibitive.