Utnyttja maskininlärningsmetoder tillsammans med data om kemisk transport, väder och markanvändning för uppskattning av organiska aerosolers komponenter

Organic aerosols (OA) are a major component of fine particulate matter and are strongly linked to elevated mortality rates. Since different OA sources have distinct impacts on human health, accurately identifying their spatial and temporal distributions is essential for effective health risk assessments and policy-making. In this thesis, we leverage a large European dataset that provides 85,000 daily OA mass concentration measurements from 175 stations, only 15,000 of which include detailed source information. While physical models (e.g., chemical transport models) can estimate OA sources, their accuracy often suffers from incomplete or uncertain emission data. To address these limitations, we incorporate outputs from these models alongside land-use and weather information as input features for a suite of Machine Learning and Deep Learning methods. Specifically, we compare Random Forest, several Gradient Boosting models, Multi-Layer Perceptrons, Gated Recurrent Units, and a beta Variational AutoEncoder, with all models fine-tuned to optimize performance and tested under different target normalization schemes. Our findings show that every Machine Learning approach outperforms CAMx—the baseline physical model used in our study. By improving source estimations at sites lacking detailed data, this work contributes to more accurate OA modeling, supporting refined health risk analyses and more effective policy interventions.