Can Deep Learning Improve CMAQ Performance?
Three-dimensional Eulerian chemical transport models such as CMAQ often report a significant model-measurement error due to uncertainties in the treatment of physical processes and require higher run-time. Machine models are more computationally efficient and are currently used widely for forecasting purposes. Deep Neural network (DNN) techniques comprise a popular class of machine learning methods. Predicting hourly air quality, especially ozone, is challenging due to its highly varying and complex behavior in the atmosphere. Here, we used modeled meteorological parameters (by MCIP) along with selected modeled gaseous species (by CMAQ) as our inputs for predicting future ozone concentrations. A timely-efficient 1D deep convolutional neural network (ConvNet 1D), called CMAQ-CNN, was implemented and trained on using CMAQ outputs as inputs to predict hourly ozone concentration in real-time across the continental US (1081 AQS stations in 48 states). The CMAQ-CNN model significantly improved the performance of the CMAQ model in term of both accuracy (IOA) and bias (maximum daily ozone). IOA improved around 0.06 in average and up to 0.3 across the United States by using CMAQ-CNN model. The CMAQ-CNN model shows mediocre performance on capturing very high ozone peaks (over 90 ppb). This poster was presented at the Earth Science Information Partners (ESIP) Winter Meeting in January 2019.