Multisource Observation-Constrained Tuning for Air Quality Forecasting in a Machine Learning-Enhanced NOAA Unified Forecast System

Abstract

Accurate near-real-time (NRT) air quality forecasting is critical for protecting public health but remains challenged by uncertainties in emissions, meteorology, and initial conditions within traditional chemical transport models (CTMs). We present an observation-driven machine learning framework that integrates multisource satellite products, including VIIRS aerosol optical depth and TEMPO NO2 with U.S. EPA AirNow measurements into a machine-learning surrogate of the NOAA Unified Forecast System (DeepAQM). A ConvLSTM-ResNet fusion module reconstructs spatially continuous surface PM2.5 and O3 fields from sparse observations, providing updated initial conditions for 72 h forecasts across the continental United States. DeepAQM reproduces UFS-AQM spatial patterns with strong agreement (R2 > 0.9 for O3 and up to 0.80 for PM2.5) while maintaining normalized mean bias within ±10%. Observation-updated initial conditions improve early forecast skill, increasing site-level R2 during the first 10 forecast hours from approximately 0.10–0.15 to 0.40–0.50 for PM2.5 and from approximately 0.40 to 0.50 for O3. Observation-based fine-tuning provides additional improvements throughout the 72 h forecast period, increasing overall R2 to 0.41 for PM2.5 and 0.82 for O3 during the independent evaluation period. These results demonstrate that integrating satellite and ground observations within DeepAQM can enhance NRT air quality forecasting and provide a scalable pathway toward globally deployable observation-driven prediction systems.