Offline Meteorology-Pollution Coupling Global Air Pollution Forecasting Model with Bilinear Pooling
Xu Fan, Yuetan Lin, Bing Gong, Hao Li · 2025
Read Offline Meteorology-Pollution Coupling Global Air Pollution Forecasting Model with Bilinear Pooling by Xu Fan, Yuetan Lin, Bing Gong, Hao Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-18405.txt.
Introduction
Fan, Lin, Gong, and Li built this air pollution forecaster in twenty twenty-five. Physics models that couple weather and pollution are accurate but slow. Learning-based rivals often weld the two together during training, which demands huge resources. The team instead couples them offline, after separate training, using a pooling method that mixes the two streams.
Their model needs only a small fraction of the parameters of online rivals. Against the leading global forecast system, it won on most variables at most lead times. The edge grew for forecasts beyond two days. Adding weather fields cut overall error by about fifteen percent. The authors present offline coupling as a lean path to real-time global warnings.
Catalog record
| CREATOR | Xu Fan, Yuetan Lin, Bing Gong, Hao Li |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-18405
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Offline Meteorology-Pollution Coupling Global Air Pollution Forecasting Model with Bilinear Pooling Xu Fan, Yuetan Lin, Bing Gong, Hao Li Abstract Air pollution has become a major threat to human health, making accurate forecasting crucial for pollution control. Traditional physics-based models forecast global air pollution by coupling meteorology and pollution processes, using either online or offline methods depending on whether fully integrated with meteorological models and run simultaneously. However, the high computational demands of both methods severely limit real-time prediction efficiency. Existing deep learning (DL) solutions employ online coupling strategies for global air pollution forecasting, which finetune pollution forecasting based on pretrained atmospheric models, requiring substantial training resources. This study pioneers a DL-based offline coupling framework that utilizes bilinear pooling to achieve offline coupling between meteorological fields and pollutants. The proposed model requires only 13% of the parameters of DL-based online coupling models while achieving competitive performance. Compared with the state-of-the-art global air pollution forecasting model CAMS, our approach demonstrates superiority in 63% variables across all forecast time steps and 85% variables in predictions exceeding 48 hours. This work pioneers experimental validation of the effectiveness of meteorological fields in DL-based global air pollution forecasting, demonstrating that offline coupling meteorological fields with pollutants can achieve a 15% relative reduction in RMSE across all pollution variables. The research establishes a new paradigm for real-time global air pollution warning systems and delivers critical technical support for developing more efficient and comprehensive AI-powered global atmospheric forecasting frameworks. 1 Introduction The rapid expansion of human productivity, coupled with accelerated industrialization, urbanization, and the global dependence on fossil fuels, has led to a significant increase in pollutant emissions [ 28 ] . These emissions undergo complex chemical reactions and atmospheric transport processes, leading to severe air pollution, which causes approximately 6.7 million premature deaths annually [ 27 ] . According to the World Health Organization (WHO), major pollutants in the atmosphere include particulate matter (PM) with diameters less than 10 micrometers (PM 10 ) and 2.5 micrometers (PM 2.5 ), sulfur dioxide (SO 2 ), carbon monoxide (CO), ozone (O 3 ), and nitrogen dioxide (NO 2 ) [ 27 ] . In response to this public health and environmental crisis, governments worldwide implement proactive measures through development and enforcement of stringent environmental regulations [ 17 , 4 ] . Advancements in global air pollution forecasting systems enhance the capacity to provide early warnings for pollution events, reduce exposure risks, and support regional pollution control, policy development, and climate change mitigation, contributing to sustainable development [ 24 ] . Air pollution forecasting is a critical part of the global atmospheric forecasting system, where meteorological fields play a crucial role in the formation, transport, and transformation of pollutants [ 19 , 2 ] . Traditional numerical forecasting methods couple meteorological forecasting with air pollutants to simulate pollutants transport dynamics, and chemical transformations on a global scale. Recent advances in DL-based weather forecasting domain have significantly improved performance of global meteorological forecasting. Models such as FourCastNet [ 21 ] , Pangu-Weather [ 3 ] , GraphCast [ 15 ] , FuXi [ 8 , 25 ] and AIFS [ 16 ] achieve competitive forecasting precision rivaling that of the high-resolution forecasts (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF), one of the world’s leading weather prediction systems. These advancements provide a strong foundation for developing an AI-powered global atmospheric forecasting system. However, research on DL-based global air pollutant forecasting remains limited. Using DL-based methods to couple air pollution forecasting with meteorological fields can overcome the constraints of traditional physical parameterization approaches, enhancing the efficiency and accuracy of air pollution predictions [ 5 ] . The meteorology-pollution coupling of traditional air pollution forecasting models can be classified into online and offline methods [ 20 ] . For online methods, the air pollution model is integrated into the meteorological model, solving the continuity equations alongside the atmospheric dynamical equations [ 13 ] . In contrast, the offline methods treat the pollution model as an independent module, using external meteorological data ( e.g . , reanalysis or numerical weather prediction data) as input to solve the continuity equations [ 7 ] . Compared to online methods, offline methods are more computationally efficient and easier to deploy. However, both methods require meteorological forecasts for coupling, further increasing the computational burden of already expensive numerical weather prediction models and necessitating substantial resources for both development and operation. DL-based models improve both forecasting efficiency and accuracy. Aurora [ 5 ] fine-tunes a foundation model pre-trained on meteorological data using the Copernicus Atmosphere Monitoring Service (CAMS) dataset [ 9 ] , the longest publicly available global air pollution dataset, enabling deep learning-based online coupled forecasting. However, its accuracy depends on the pre-training process, and online coupled fine-tuning is computationally expensive. The DL-based offline approach for global air pollution forecasting demonstrates enhanced flexibility and adaptability to diverse meteorological field inputs, while simultaneously imposing a high demand for the intergration of meteorological and pollution data. Bilinear pooling enables effective multimodal feature integration through high-order interactions that approximates vector outer products [ 18 , 10 ] . By developing a low-rank approximation scheme combined with expanded receptive fields, we achieve efficient fusion of meteorological field and pollutant features. Furthermore, the bilinear fusion inherently enables low-cost adaptation to different meteorological input configurations. In this study, we present a novel offline-coupling AI framework for global air pollution forecasting which integrates meteorological fields and pollutants. The framework employs separate processing of meteorological and pollutant data streams, and designs an offline coupling module based on bilinear pooling, which collaboratively fuses multi-modal data. This architecture enhances computational efficiency while improving adaptability to heterogeneous meteorological forecasting models and enabling flexible cross-platform deployment. Additionally, we demonstrate the role of meteorological fields in global air pollution forecasting. These advancements provide both methodological innovations and technical support for developing AI-powered atmospheric forecasting systems. The contributions can be summarized as follows: • We propose an DL-based offline meteorology-pollution coupling framework that achieves competitive results with only 13% of model parameters compared with online methods. We quantitatively validate the contribution of meteorological fields to air pollution prediction. • For meteorological and pollutant dual-modal data, we introduce deep integration via bilinear pooling, innovatively establishing nonlinear correlations between meteorological elements and pollutant dispersion in feature space. 2 Related Works Traditional physics-based models. Traditional physics-based models conduct global forecasts by solving a set of 3D continuity equations that describe emissions, transport, chemical reactions, and deposition [ 6 ] . Only a few models are capable of global air pollution forecasting. According to the World Meteorological Organization (WMO) Global to Local Air Quality Forecast Inventory [ 26 ] , the primary models offering real-time global air pollution forecasts include: the National Aeronautics and Space Administration’s (NASA) GEOS-CF system, the National Center for Atmospheric Research’s (NCAR) Community Earth System Model Version 2 (CESM2), the Finnish Meteorological Institute’s SILAM, and the CAMS system [ 9 ] . These models offer global air pollution forecasts with spatial resolutions ranging from 0.25° to 1° and forecast lead times of 4 to 10 days. In this study, CAMS’s data are selected for training and forecast performance comparison. CAMS system is operated by ECMWF, extends the atmospheric composition