Zeeman: A Deep Learning Regional Atmospheric Chemistry Transport Model

TL;DR

Zeeman introduces a boundary-aware, attention-based deep learning framework to perform fast, multi-pollutant regional atmospheric chemistry forecasting by surrogate modeling LOTOS-EUROS outputs. The architecture uses boundary enhancement, a 3D cube embedding, and Swin Transformer blocks to predict $O_3$, $NH_3$, $NO_2$, $PM_{2.5}$, and $PM_{10}$ while incorporating hourly emissions and ECMWF meteorology in an auto-regressive setup. Key findings show Zeeman achieving near-parallel performance to a state-of-the-art CTM across spatial, vertical, and 5-day horizons, with strong ozone and particulate matter forecasts and diurnal variability tied to emissions and meteorology. The work highlights potential for data assimilation, emission inversion, and offline coupling with weather models, enabling efficient, accurate regional air-quality forecasts at orders-of-magnitude faster speeds than traditional CTMs.

Abstract

Atmospheric chemistry encapsulates the emission of various pollutants, the complex chemistry reactions, and the meteorology dominant transport, which form a dynamic system that governs air quality. While deep learning (DL) models have shown promise in capturing intricate patterns for forecasting individual atmospheric component - such as PM2.5 and ozone - the critical interactions among multiple pollutants and the combined influence of emissions and meteorology are often overlook. This study introduces an advanced DL-based atmospheric chemistry transport model Zeeman for multi-component atmospheric chemistry simulation. Leveraging an attention mechanism, our model effectively captures the nuanced relationships among these constituents. Performance metrics demonstrate that our approach rivals numerical models, offering an efficient solution for atmospheric chemistry. In the future, this model could be further integrated with data assimilation techniques to facilitate efficient and accurate atmospheric emission estimation and concentration forecast.

Figures (24)

• Figure 1: Overall architecture of Zeeman. (a) the main components of Zeeman: boundary enhancement, cube embedding, Swin-transformer, full connected layer. (b) the process of cyclic prediction based on Zeeman.
• Figure 2: Spatial distribution of correlation coefficient (R) from Zeeman prediction and LOTOS-EUROS simulations for variables on ground level with the lead time of 6, 12, 18, 24 hours. All the predictions start from 00:00 at each day on 2022.
• Figure 3: Vertical distribution of correlation coefficient (R) between Zeeman forecast and LOTOS-EUROS with the lead time of 6, 12, 18, 24 hours. All the forecasts start from 00:00 at each day on 2022.
• Figure 4: Trend of correlation coefficient (R) for the hourly forecasts made by Zeeman. The lead time starts from 1 hour to 120 hour. The shown data is on ground level and all the forecasts start from 00:00 at each day on 2022.
• Figure 5: Time series of 5 day forecast on Rotterdam (red), Groningen (blue) and Düsseldorf (green). Dash line is LOTOS-EUROS simulations and solid line is Zeeman forecast. The forecast starts from 00:00, 1st June, 2022.