A LASSO-based reduced-form CMAQ model for predicting ozone and PM2.5 responses to emission changes in South Korea.

Abstract

Reduced-form models of the Community Multiscale Air Quality Modeling System (CMAQ) enable efficient prediction of air quality responses to emission changes. In this study, we developed a reduced-form CMAQ model based on the least absolute shrinkage and selection operator (LASSO) to approximate CMAQ outputs in high-dimensional settings where the number of emission variables exceeds the number of training samples. CMAQ simulations were generated using 118 emission scenarios covering seven emission sectors across 17 provincial-level administrative regions in South Korea. To account for the bounded nature of pollutant concentrations and to better represent nonlinear responses, an adaptive logit transformation was applied within the LASSO framework. The model was trained on 100 simulations and evaluated on 18 test scenarios, achieving mean root mean square errors of 0.1 ppb for ozone and 0.1 μg/m3 for PM2.5. The model also identifies a small subset of influential sector-region emission variables, enabling interpretable analysis of emission impacts and supporting the design of targeted emission scenarios. A web-based interface was developed to demonstrate the applicability of the approach, allowing interactive exploration of pollutant responses to emission changes. The reduced-form model enables rapid prediction of air quality responses with substantially lower computational cost than CMAQ simulations.