Deep Learning Ensembles for Improved Atmospheric Composition Modeling

Abstract

Better forecasting of atmospheric composition is a critical aspect of environmental and climate monitoring. Among climate and weather numeric modeling, often ensembles are used to improve the forecasting power and to quantify the uncertainty of the model. However, the numerical simulation of atmospheric chemistry, critical for composition simulations, is computationally too expensive to generate numerical composition ensembles. One way to address this problem is to use deep learning to emulate the slow physical model. In this work we study the feasibility of two different deep learning methods and show how an emulator could be used to realistically estimate uncertainties of atmospheric composition forecasts, bypassing the need to run costly numerical ensemble simulations. One of the methods builds upon Fourier neural operators and the NVIDIA FourCastNet architecture and the second method builds on conditional Generative Adversarial Networks. We design the models to respond to perturbations to the most important drivers of air pollution, including meteorology and pollutant emissions. We apply this framework to the NASA GEOS Composition Forecast System (GEOS-CF), which produces daily global composition forecasts at approximately 25 km^2 horizontal resolution. Due to computational constraints, GEOS-CF currently has limited capability to produce probabilistic estimates or to optimally assimilate trace gas observations. We show how a deep learning emulator has the potential to improve composition forecasts produced by GEOS-CF or other, similar types of applications. These methods could be applied to other types of ensemble-based models, potentially providing a large speed-up in overall modeling time.