Emulating CMAQ Simulations with a Conditional U-Net: A Data-Driven Approach for PM2.5 Concentration Prediction

Abstract

Fine particulate matter (PM2.5) pollution poses significant health risks, necessitating accurate predictions for effective management measures for the mitigator of PM2.5. While the Community Multiscale Air Quality (CMAQ) model is widely utilized to simulate PM2.5 concentrations, its computational demands often limit its application in testing various emission reduction scenarios. This study introduces a data-driven approach to emulate CMAQ simulations through AI surrogate model with a conditional U-Net architecture. The model, trained on a dataset of CMAQ-simulated PM2.5 concentrations from across South Korea in 2013, incorporates region-specific emission source activity data as inputs. It accurately predicts hourly changes in PM2.5 concentrations, facilitating the estimation of monthly and annual average concentrations under diverse emission reduction scenarios. Demonstrating robust performance in replicating the spatial and temporal patterns of CMAQ simulations, this surrogate model significantly cuts computational costs and times. Our findings offer policymakers a practical and efficient tool for evaluating the effectiveness of emission control strategies and enhancing decision-making processes in air quality management.