Ozone response modeling to NOx and VOC emissions: Examining machine learning models

Area of Science:

• Atmospheric chemistry and air quality modeling.
• Application of machine learning in environmental science.

Background:

• Current machine learning (ML) in atmospheric science primarily focuses on forecasting and bias correction.
• Few studies have investigated the nonlinear responses of ML predictions to precursor emissions.
• Understanding ozone (O3) response to emissions is crucial for air quality management.

Purpose of the Study:

• To examine O3 responses to local anthropogenic nitrogen oxides (NOx) and volatile organic compound (VOC) emissions in Taiwan using Response Surface Modeling (RSM).
• To compare the performance of different ML approaches against numerical modeling.
• To assess the interpretability and potential biases of ML predictions in air quality analysis.

Main Methods:

• Utilized ground-level maximum daily 8-hour ozone average (MDA8 O3) as a key indicator.
• Employed Response Surface Modeling (RSM) to analyze O3 responses to NOx and VOC emissions.
• Compared three datasets: Community Multiscale Air Quality (CMAQ) model data, ML-measurement-model fusion (ML-MMF) data, and ML data.

Main Results:

• ML-MMF (r = 0.93-0.94) and ML (r = 0.89-0.94) predictions significantly outperformed CMAQ predictions (r = 0.41-0.80).
• ML-MMF isopleths accurately reflected O3 nonlinearity due to numerical and observational corrections.
• ML-only predictions showed biases and distorted O3 responses, potentially misleading air quality control targets.
• Transboundary pollution from mainland China impacts regional O3 sensitivity to local emissions.

Conclusions:

• ML-MMF offers a more reliable approach for air quality prediction and policy assessment than ML alone.
• Interpretability and explainability are crucial for future ML applications in atmospheric science.
• Combining ML with physical/chemical mechanisms and robust statistical methods is essential for accurate air quality modeling.