Explainable AI for predicting oxidative potential of fine particles and key chemical drivers

Abstract

Oxidative potential (OP) has emerged as promising health metric for ambient fine particles. Chemical components and OP of fine particles measured in China and Korea were used to develop OP prediction model with understanding the influence of chemical components and their interaction. Mn, Cu, Zn, Pb, and water-soluble organic carbon (WSOC) were selected as key chemical components to affect the OP. Various machine learning models incorporating explainable AI techniques were trained and evaluated. The best prediction model was found to be voting regression which aggregated individual predictions from random forest and gradient boosting models, explaining 74.9 % of OP variabilities across all measurement sites. Mn was the most important feature to affect the OP, followed by Pb, WSOC, Cu, and Zn. During OP event days at urban Gwangju, the Pb became the most important contributor, while at agricultural Gimje, the WSOC was the one to affect the OP. It was also found that the Cu above 0.004 µg/m³ with the WSOC had a strong antagonistic effect on the OP. The explainable AI methods should be so useful to predict the OP of ambient fine particles and to understand the important chemical components and their interaction. © 2025 Elsevier B.V., All rights reserved.