AGU Fall Meeting 2021

The Integration of Artificial Intelligence for Improved Operational Air Quality Forecasting

, , , , , , , and

The National Oceanic and Atmospheric Administration (NOAA) is actively seeking to integrate the latest research in Artificial Intelligence (AI) techniques to solve a number of operational challenges. We describe two efforts underway that integrate deep learning techniques to improve operational air quality (AQ) forecasting. The first effort is a collaboration between NOAA and the University of Maryland Baltimore County (UMBC) that uses deep learning to improve bias correction of model predicted ozone and fine particulate matter (PM 2.5) concentrations. Bias correction is typically applied as a post-processing method to correct model errors. The current approach combining Kalman filter and analog methods has a long history of success, and it is challenged by extreme events, which are increasing in frequency. This research explores the integration of a deep learning algorithm, trained to learn historical patterns of extreme events, into the bias correction process to augment the existing methods for improved bias correction with emphasis on extreme event detection. The second effort is a collaboration among NOAA, UMBC, Johns Hopkins University (UCAR/JHU/APL) and USRA, exploring the use of machine learning emulators for accelerated transport of atmospheric tracers: chemical species and aerosols. NOAA provides operational AQ forecast guidance, including ozone and PM 2.5, based on the Community Multiscale Air Quality (CMAQ) Modeling System at 12 km resolution. In support of NOAA’s current initiative to develop the Unified Forecast System and the next generation AQ forecasting system, featuring online coupling of CMAQ with the Rapid Refresh Forecast System (RRFS), machine learning is being introduced to reduce the computational costs of the transport of chemical and aerosol tracers, which account for 40% of the overall computation. This effort supports representing chemical transformations at increasingly finer spatial resolutions, with the goal of 3 km resolution, which would otherwise be intractable. This study includes an evaluation of deep convolutional neural nets for spatio-temporal learning of nonlinear dynamic transport and is part of a larger collaboration with NASA to include machine learning emulation of both the transport and the chemical processes for atmospheric composition predictions.


UMBC ebiquity