Constructing transferable and interpretable machine learning models for black carbon concentrations.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Black carbon (BC) poses health risks but lacks widespread in-situ monitoring.
• Machine learning (ML) offers a solution as virtual sensors for air quality data.
• This study focuses on the transferability and interpretability of ML models for BC estimation.

Purpose of the Study:

• To evaluate and compare white-box (WB) and black-box (BB) ML models for estimating BC concentrations.
• To assess the transferability of ML models trained in one urban background site to other European urban and traffic sites.
• To analyze the interpretability of different ML models in quantifying feature importance for BC estimation.

Main Methods:

• Trained and tested multiple WB and BB ML models using long-term air pollutant and weather data from Barcelona.
• Evaluated model performance in Helsinki (traffic) and Dresden (urban background) sites.
• Employed various interpretation techniques to quantify feature importance for each model.

Main Results:

• Black carbon (BC) strongly correlates with particle number concentration of accumulation mode (PNacc) and nitrogen dioxide (NO2).
• ML models trained in Barcelona demonstrated strong performance in Helsinki and Dresden, indicating good transferability.
• Black-box models, particularly the long short-term memory (LSTM) model, generally outperformed white-box models in explaining BC variability.

Conclusions:

• Interpretable ML models can be effectively transferred across different geographical locations and site types (urban background, traffic).
• PNacc and NO2 are key predictors, but their influence can be positive or negative depending on the site.
• The study highlights the potential of transferable and interpretable ML models for enhancing air quality monitoring networks.