Difference Learning for Air Quality Forecasting Transport Emulation

TL;DR

The paper tackles the computational bottleneck in NOAA's UFS-AQ transport module, which hinders achieving 3 km air-quality forecasts. It introduces a difference-learning DL surrogate based on a 3D U-Net to predict per-timestep advection residuals, exploiting the advection invariance and learning a per-patch normalization that is species-agnostic. On NOAA AQM.v7.0 runs, the model achieves an overall RMSE of $0.0115$ in min-max space, maintains low errors during extreme events, and infers a full CONUS timestep in about $2.6$ seconds, with a mass-conservation metric showing a $0.0741\%$ mean difference for a PM2.5 constituent. These results indicate the potential for operational deployment of higher-resolution AQ guidance, with future work focusing on data transformations, larger overlapping patches, and physics-informed regularization to enforce mass conservation.

Abstract

Human health is negatively impacted by poor air quality including increased risk for respiratory and cardiovascular disease. Due to a recent increase in extreme air quality events, both globally and locally in the United States, finer resolution air quality forecasting guidance is needed to effectively adapt to these events. The National Oceanic and Atmospheric Administration provides air quality forecasting guidance for the Continental United States. Their air quality forecasting model is based on a 15 km spatial resolution; however, the goal is to reach a three km spatial resolution. This is currently not feasible due in part to prohibitive computational requirements for modeling the transport of chemical species. In this work, we describe a deep learning transport emulator that is able to reduce computations while maintaining skill comparable with the existing numerical model. We show how this method maintains skill in the presence of extreme air quality events, making it a potential candidate for operational use. We also explore evaluating how well this model maintains the physical properties of the modeled transport for a given set of species.

Figures (4)

• Figure 1: Distributions of chemical species anai, a PM2.5 constituent, and ozone at vertical levels 1 (blue) and 16 (red). The $Input$ and $Output$ columns are concentration distributions from the UFS-AQ model after min-max normalization. Column 3 is the distribution of the difference between the $Input$ and $Output$ data. Column 4 is the distribution after taking the cube root of this difference. Note that the y-axis is log-scaled.
• Figure 2: U-Net predictions, where $Ground \; Truth = (Output - Input)^{1/3}$. The first three columns show patches from vertical layer 1, and the last three columns show patches from vertical layer 16. The first two rows show non-extreme patches, and the last two rows show extreme patches. Species asvpo1j and aclk are PM2.5 constituents, and ald2 and hno3 are contributors to ozone and PM2.5 formation respectively.
• Figure 3: Illustration of the difference learning approach. The U-Net learns the mapping between $Input$ and $(Output - Input)^{1/3}$. The $Predicted \; Output$ is the U-Net's prediction in the min-max normalized space.
• Figure 4: Effects of $(Output - Input)^{1/n}$ transformations on the distribution of asvpo2i, a PM2.5 constituent, at vertical levels 1 (blue) and 16 (red). From left to right, top to bottom, $n = 1, 3, 5, 7, 9, 15$.