Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diffusion-based Probabilistic Air Quality Forecasting with Mechanistic Insight

Ao Ding, Aoxing Zhang, Tzung-May Fu, Yuanlong Huang, Qianjie Chen, Yuyang Chen, Jiajia Mo, Wei Tao, Wai-Chi Cheng, Lei Zhu, Xin Yang, Guy Brasseur

Abstract

Current operational air quality forecasts are computationally expensive, sensitive to errors in physics and emissions, and often neglect weather-related uncertainty. To address these limitations, we present AirFusion, a hybrid, diffusion-based framework that synergistically integrates knowledge from chemical transport models with real-world observational constraints to enable accurate and efficient probabilistic regional air quality prediction. We apply AirFusion to generate operational 6-day, 30-member ensemble forecasts of surface ozone across China, initialized with observations and driven by ensemble weather forecasts. AirFusion outperforms existing operational benchmarks, achieving substantially lower forecast errors against surface measurements, while also providing ensemble-based diagnostics that explicitly quantify the impacts of weather uncertainty on air quality predictability. Moreover, AirFusion can rapidly adapt to evolving emissions through fine-tuning with only one month of recent observations. These attributes establish AirFusion as a powerful and extensible framework for next-generation probabilistic air quality forecasting, with clear potential for application to other pollutants and regions.

Diffusion-based Probabilistic Air Quality Forecasting with Mechanistic Insight

Abstract

Current operational air quality forecasts are computationally expensive, sensitive to errors in physics and emissions, and often neglect weather-related uncertainty. To address these limitations, we present AirFusion, a hybrid, diffusion-based framework that synergistically integrates knowledge from chemical transport models with real-world observational constraints to enable accurate and efficient probabilistic regional air quality prediction. We apply AirFusion to generate operational 6-day, 30-member ensemble forecasts of surface ozone across China, initialized with observations and driven by ensemble weather forecasts. AirFusion outperforms existing operational benchmarks, achieving substantially lower forecast errors against surface measurements, while also providing ensemble-based diagnostics that explicitly quantify the impacts of weather uncertainty on air quality predictability. Moreover, AirFusion can rapidly adapt to evolving emissions through fine-tuning with only one month of recent observations. These attributes establish AirFusion as a powerful and extensible framework for next-generation probabilistic air quality forecasting, with clear potential for application to other pollutants and regions.
Paper Structure (11 sections, 5 figures)

This paper contains 11 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Overview of the AirFusion ensemble air quality forecast system
  3. Application of AirFusion to forecast surface ozone over China
  4. Multi-scale feature learning of AirFusion
  5. Fast and flexible adaptation of AirFusion to changes in pollutant emissions
  6. Probabilistic air quality forecasting with AirFusion
  7. Discussion
  8. Methods
  9. Regional ozone simulation using the WRF-GC model
  10. Surface ozone measurements, down-scaled meteorological reanalysis, and ensemble weather forecasts over China
  11. Construction and training of the AirFusion system

Figures (5)

  • Figure 1: Schematic of the AirFusion ensemble air quality forecast system.a,b,e, Training of AirFusion's three modules: AirFusion-S (a), AirFusion-T (b), and AirFusion-T-FT (e). c, Given surface observations of pollutant concentrations, AirFusion-S generates a continuous 2-D pollutant concentration field, $X^0(t)$. e, In operational forecast mode, AirFusion-S generates initial conditions, $X^0(t_0-\Delta t)$ and $X^0(t_0)$, from surface pollutant observations. Weather forecasts are then input to AirFusion-T-FT to predict surface pollutant concentrations at the next time step, $X^F(t_0+\Delta t)$. Subsequent forecasts use the AirFusion-T-FT outputs and ensemble weather forecasts to advance in time through the desired forecast period. The above procedure is then repeated for all $N$ members of an ensemble weather forecast.
  • Figure 2: Performance of the AirFusion system in forecasting surface ozone concentrations across China.a-c, Root mean square errors (RMSEs) of Day-1 MDA8O3 forecasts for July to September 2024 from three models: WRF-GC (a), AirFusion without fine-tuning (AirFusion-noFT) (b), and the fully fine-tuned AirFusion system (c). MDA8O3 forecasts in AirFusion-noFT and AirFusion were derived by linearly interpolating their respective 3-hourly ensemble mean ozone forecasts. d, Comparisons of RMSEs of MDA8O3 (solid bars) and 3-hourly ozone (hatched bars) concentrations averaged across 341 Chinese cities for Day 1 to Day 6 lead times. Green, WRF-GC; orange, AirFusion-noFT; purple, AirFusion; blue, AirFusion-nudgedH hindcast. Black whiskers indicate standard deviations across Chinese cities. Also shown are the RMSEs over four major megacity clusters (colored symbols): Beijing–Tianjin–Hebei (BTH), Yangtze River Delta (YRD), Pearl River Delta (PRD), and Sichuan–Chongqing (CY), along with the range of RMSEs from previous multi-model assessments over the BTH zhu2024comparative (black diamond and whiskers) and ensemble model over eastern Chinapetersen2019ensemble (grey shaded area).
  • Figure 3: Spatial scales of surface ozone pollution features learned by the AirFusion system.a-c, Two-dimensional discrete wavelet decomposition of forecasted MDA8O3 concentration differences between WRF-GC and AirFusion-noFT on Day 1 at scales $\ge$432 km (a) and <432 km (b), and the energy distributions on Days 1 (outer) and 3 (inner) (c). d-f, Two-dimensional discrete wavelet decomposition of forecasted MDA8O3 concentration differences between AirFusion and AirFusion-noFT on Day 1 at scales $\ge$432 km (d) and <432 km (f), and the energy distributions on Days 1 (outer), 3 (mid), and 5 (inner) (f).
  • Figure 4: Comparison of AirFusion's performance in predicting surface ozone of August 2020 when fine-tuned with observations in June 2024 (blue), June 2020 (orange), and July 2020 (green). a-c, RMSEs (a), temporal correlations (b), and NMBs (c) for Day 1 to Day 6. Solid lines and shaded areas indicate the national means and standard deviations across 341 Chinese cities, respectively. d, mean RMSEs for Days 1 to 6.
  • Figure 5: Probabilistic ozone forecasting with AirFusion. a, Time series of observed (black) and ensemble-mean forecast MDA8O3 concentrations for Shenzhen during July–September 2024. The dark- and light-blue lines show Day-1 and Day-5 forecasts, respectively; shading indicates the standard deviation across ensemble members. RMSEs and standard deviations of the ensemble mean against observations are shown inset. b, Comparison of the ozone exceedance probability (OEP) forecast by AirFusion with the observed exceedance frequency across 341 Chinese cities during July–September 2024, using the WHO guideline (100 $\mu$g m$^{-3}$ for MDA8O3) as the threshold.