Machine learning-based downscaling of aerosol size distributions from a global climate model

Abstract

Air pollution, particularly exposure to ultrafine particles (UFPs) with diameters below 100 nm, poses significant health risks, yet their spatial and temporal variability complicates impact assessments. This study explores the potential of machine learning (ML) techniques in enhancing the accuracy of a global aerosol-climate model’s outputs through statistical downscaling to better represent observed data at specific sites. Specifically, the study focuses on the particle number size distributions from the global aerosol-climate model ECHAM-HAMMOZ. The coarse horizontal resolution of ECHAM-HAMMOZ (approx. 200 km) makes modeling sub-gridscale phenomena, such as UFP concentrations, highly challenging. Data from three European measurement stations (Helsinki, Leipzig, and Melpitz) were used as target of downscaling, covering nucleation, Aitken, and accumulation particle size ranges during years 2016–2018. Six different ML methods (Random Forest, XGBoost, Neural Networks, Support Vector Machine, Gaussian Process Regression and Generalized Linear Model) were employed, with hyperparameter optimization and feature selection integrated for model improvement. A separate ML model was trained for each of the sites and size ranges. Results showed a notable improvement in prediction accuracy for all particle sizes compared to the original global model outputs, particularly for the accumulation subrange. Challenges remained particularly in downscaling the nucleation subrange, likely due to its high variability and the discrepancy in spatial scale between the climate model representation and the underlying processes. Additionally, the study revealed that the choice of downscaling method requires careful consideration of spatial and temporal dimensions as well as the characteristics of the target variable, as different particle size ranges or variables in other studies may necessitate tailored approaches. The study demonstrates the feasibility of ML-based downscaling for enhancing air quality assessments. This approach could support future epidemiological studies and inform policies on pollutant exposure. Future integration of ML models dynamically into global climate model frameworks could further refine climate predictions and health impact studies.