Knowledge-informed deep learning to mitigate bias in joint air pollutant prediction.

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Computational Science

Background:

• Accurate air pollutant prediction is vital for public health and environmental management.
• Traditional machine learning (ML) methods lack physicochemical constraints, causing biases in exposure estimates.
• Chemical transport models struggle with local-scale variability crucial for health assessments.

Purpose of the Study:

• To develop a novel physics-informed deep learning framework for high-resolution, bias-free air pollutant prediction.
• To integrate atmospheric physics, including advection-diffusion and fluid dynamics, into neural networks.
• To improve air quality modeling for epidemiological studies and health impact assessments.

Main Methods:

• Developed a physics-informed deep learning framework incorporating advection-diffusion equations and fluid dynamics constraints.
• Modeled air pollutant pairs (NO2/NOx in California; PM2.5/PM10 in China) using integrated physical constraints.
• Employed proxy advection and diffusion fields to alter learning dynamics and reduce generalization error.

Main Results:

• Achieved significant bias reduction: 21%-42% for nitrogen oxides and 16%-17% for particulate matter compared to baseline ML methods.
• Demonstrated improved computational efficiency over graph networks.
• Generated physically interpretable parameters and provided uncertainty quantification via ensemble techniques.

Conclusions:

• The physics-informed deep learning framework offers a high-resolution, physically constrained approach to air quality modeling.
• This methodology substantially reduces systematic bias in pollutant predictions, enhancing reliability for health studies.
• The framework's interpretability and uncertainty quantification advance air pollutant exposure assessment and epidemiological research.