Deep Learning Ensemble Emulation for NASA’s Goddard Earth Observing System (GEOS) Composition Forecast

Abstract As poor air quality continues to pose threats to humanity, better modeling of atmospheric composition is critical both for environmental and climate monitoring. The NASA Goddard Earth Observing System Composition Forecast (GEOS-CF) provides daily global air quality forecasts of atmospheric composition. Due to the full chemistry mechanisms and transport of hundreds of chemical tracers, GEOS-CF forecasts cannot extend beyond 5 days, and probabilistic estimates of tracer concentrations are computationally infeasible to calculate. Probabilistic estimates are critical in quantifying the forecast uncertainty and are achieved by generating an ensemble of models with slightly different initializations. We describe a 2-yr study to build a deep learning ensemble emulator. In this study, we describe why models such as NVIDIA’s FourCastNet are inadequate for learning atmospheric composition and can experience instabilities during training. We share competitive forecast skill for carbon monoxide, nitric oxide, nitrogen dioxide, and ozone, using a novel air quality conditional generative adversarial network (AQcGAN) built for emulating ensembles. Using this model, we are able to forecast ten days into the future generating both ensemble mean and spread in seconds for a given ensemble, offering a low-cost option for extending the forecast window in GEOS-CF and for quantifying uncertainty. Results show that when compared to the actual mean and spread of the ensembles, the emulator forecasts offer competitive skill in terms of low RMSE scores. We share results from an operational data assimilation study where the AQcGAN has been used as the ensemble. We describe future plans for integration with data assimilation methods and GEOS-CF. Significance Statement This study explores the use of deep learning, a subfield of artificial intelligence, to emulate the numerical ensembles of the NASA Goddard Earth Observing System (GEOS) composition forecast system to provide a way to measure uncertainty and extend the forecast window from 5 to 10 days with sustained skill. The key findings include competitive results when using a deep learning emulator trained and evaluated using a 32-member numeric ensemble for a 10-day forecast of their respective mean and spread. Additionally, in a limited data assimilation experiment that used the deep learning ensemble emulator in lieu of the numeric ensemble, the forecasted results were on the same scale as the numeric ensemble. The added benefit of this approach as opposed to commercial deep learning models is its efficiency in terms of both the system resource consumption and the total training time. We continue to improve the ensemble skill and show in future work generalizability across different ensembles for both GEOS Composition Forecast (GEOS-CF) and more broadly across other NASA-related systems.