Generative Deep Learning Models for Air Quality Numeric Model Emulation

Accurate and timely forecasting of the physical and chemical processes impacting atmospheric

compounds is critical for prediction of air quality that is used as a basis for air quality forecasts

that are issued to protect public health. Current methods rely on a series of numerical

simulations that model the chemical reactions and physical transport of chemical gaseous and

aerosol species throughout the atmosphere; however, they are computationally expensive to

be operationally viable for finer-resolution forecasting. We are exploring the use of generative

deep learning models to emulate the full series of numerical simulations. We are training a

conditional generative adversarial network (cGAN) to emulate the NOAA Community Multiscale

Air Quality (CMAQ) Modeling System forecasts of species related to ozone and fine particulate

matter (PM2.5). We are evaluating the trained cGAN with a focus on its prediction accuracy,

long-term forecasting capability, and inference speed.