[Paper Review] Diffusion-based Probabilistic Air Quality Forecasting with Mechanistic Insight

Ao Ding, Aoxing Zhang+10 more|arXiv (Cornell University)|Mar 22, 2026

Atmospheric chemistry and aerosols0 citations

TL;DR

AirFusion is a hybrid diffusion-based framework that combines CTM-derived mechanistic insight with observation-driven fine-tuning to produce 6-day, 30-member ensemble surface ozone forecasts over China, delivering improved accuracy and explicit weather-uncertainty diagnostics.

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.

Motivation & Objective

• Address the limitations of conventional CTM-based forecasts (cost, physics errors, weather uncertainty) by integrating mechanistic CTM knowledge with observational constraints.
• Develop a diffusion-based framework (AirFusion) that pre-trains on CTM simulations and fine-tunes with recent observations for accuracy and adaptability.
• Enable operational, ensemble-based probabilistic forecasting to quantify meteorological uncertainty impacts on air quality.
• Demonstrate rapid adaptation to changing emissions through targeted fine-tuning with limited recent data.

Proposed method

• Use a diffusion-model backbone to generate continuous 2-D ozone concentration fields.
• Three modular components (AirFusion-S, AirFusion-T, AirFusion-T-FT) share the diffusion framework and learn from CTM simulations (WRF-GC) and observations.
• AirFusion-S interpolates sparse observations to create current-time concentration fields.
• AirFusion-T advances concentrations using present and future meteorology to the next time step.
• AirFusion-T-FT fine-tunes AirFusion-T with recent observations and forecasts to correct CTM biases and align with current emissions.
• Forecast mode propagates through time for N ensemble members driven by ensemble weather forecasts, yielding a probabilistic forecast and an ensemble spread.

Experimental results

Research questions

• RQ1 Can a diffusion-based hybrid model outperform traditional CTMs and prior AI-AQ models in 6-day ozone forecasting over China?
• RQ2 How effectively can CTM-derived mechanistic knowledge be combined with observational fine-tuning to improve accuracy and adaptability to emission changes?
• RQ3 What is the role of meteorological uncertainty in ozone forecast errors, and how can probabilistic diagnostics quantify this impact?
• RQ4 How rapidly can AirFusion adapt to changes in emissions through limited recent observations?
• RQ5 What is the operational performance and computational efficiency of AirFusion compared to CTMs?

Key findings

• AirFusion with fine-tuning (AirFusion) vastly outperforms WRF-GC and AirFusion-noFT in Day 1 RMSE and correlations across 341 Chinese cities.
• Day 1 RMSE for AirFusion: 26.9 ± 5.7 μg m-3; Day 6 RMSE: 32.8 μg m-3, with modest bias and high temporal correlation (r ≈ 0.7).
• AirFusion achieves 40-second, 30-member ensemble forecasts over China on a single RTX 4090, far faster than a single-member WRF-GC run on CPUs.
• Fine-tuning with only five months of observations markedly improves forecast skill, preserving mechanistic grounding while correcting recent emission changes.
• AirFusion provides ensemble-based probabilistic forecasts of ozone exceedance (OEP) with a mean calibration error of 0.106 for Day 1, acknowledging weather-forecast uncertainty as a primary error source.