Fast and Scalable Air Quality Neural Emulator for High‐Resolution Predictions on Very Large Domains
TL;DR
While SIRANet enables near–real‐time, high‐resolution forecasting, its performance remains conditioned by the representativeness of the SIRANE training simulations and tends to underestimate the highest concentrations in traffic‐dominated hotspots due to the smoothing of sharp spatial gradients inherent to the current U‐Net architecture.
Abstract
This research explores the development and implementation of SIRANet, an advanced deep learning‐based emulator designed for fine‐scale air quality modeling. Leveraging the computational efficiency of neural networks, SIRANet builds upon the established SIRANE atmospheric pollution modeling framework. It enables rapid and accurate simulations for various air pollution scenarios across the Grand Est French region. The system relies on U‐Net neural network architectures conditioned for different pollutant emission sources, such as industrial, residential, and traffic emissions, trained with multi‐scale data inputs projected onto a fine‐resolution 25 × 25 m grid. Comprehensive validation demonstrates strong agreement between SIRANet predictions and SIRANE outputs, with significant reductions in computation time and financial costs. While SIRANet enables near–real‐time, high‐resolution forecasting, its performance remains conditioned by the representativeness of the SIRANE training simulations and tends to underestimate the highest concentrations in traffic‐dominated hotspots due to the smoothing of sharp spatial gradients inherent to the current U‐Net architecture. The deployment of SIRANet at Atmo Grand Est highlights its potential as a scalable, operational tool for high‐resolution air quality forecasts, offering new opportunities for pollution scenario analysis and public health decision support.