Prediction of urban airborne ultrafine particle number concentrations using machine learning and deep learning algorithms (ULTRAMADE)

• Applicants: Dr. Lars Gerling; Dr. Tobias Schripp
• Subject Area: Atmospheric Science ; Physical Geography
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 524386321

Ultrafine particles (UFP) with an aerodynamic diameter smaller than 100 nm are suspected to harm human health. Nonetheless, conclusive evidence from epidemiological studies is still lacking. In order to derive exposure concentrations for ambient UFP, statistical modelling approaches have been used to predict UFP number concentrations. One of the more common approach is the land-use regression (LUR) based on a linear modelling approach. However, other, more sophisticated statistical modelling approaches have been used in air quality studies, e.g. machine learning (ML) and deep learning (DL) approaches, which promise an increased prediction accuracy. The aim of the project is the modelling of urban UFP number concentrations in urban environments based on ML and DL algorithms. These algorithms promise to improve the prediction accuracy compared to linear modelling approaches. We aim at representing both spatial and temporal variation with our model approach. In a first stage, we will use results from mobile measurement campaigns for calibrating a ML-based LUR model. Further effort will be done by identifying urban emissions from local sources other than vehicle traffic and explicitly including them into the model. In a second stage, a DL modelling approach based on long-term UFP measurements will be coupled with the ML model in order to improve the representation of the temporal variation of the model. Our proposed work programme is composed of five work packages (WP): WP 1 includes mobile measurements using a mobile lab and a measurement bicycle. WP 2 consists of stationary measurements performed at German Ultrafine Aerosol Network sites. In WP 3, we will identify and quantify important UFP emission sources with additional short-term stationary measurements, especially non-traffic emissions. In WP 4, we will use ML algorithms to create a statistical model. As calibration data, we will use the measurement results from WP 1. The resulting model will predict the UFP number concentration from a set of explanatory variables, i.e. meteorological quantities, land use, urban morphology, amounts of traffic and additional information about UFP sources as gathered from WP 3. In WP 5, we will use the UFP number concentrations from WP 2 for a DL modelling approach which will be able to represent the temporal variation. This approach will be coupled with the ML model from WP 4. The benefit from the model coupling will be validated using the data set generated in WP 3. We hypothesize that the project will result in a modelling framework that will be able to represent both spatial and temporal variation of urban UFP number concentrations in a high accuracy. Therefore, our study will contribute further to the quantification of exposure concentrations to ambient UFP in epidemiological research.

• DFG Programme: Research Grants