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

Authors

• Lianfa Li, Chinese Academy of SciencesFollow
• Roxana Khalili, University of Southern California
• Frederick Lurmann, Sonoma Technology, Inc.
• Nathan Pavlovic, Sonoma Technology, Inc.
• Jun Wu, University of California
• Yan Xu, University of Southern California
• Yisi Liu, University of KentuckyFollow
• Karl O’Sharkey, University of Southern California
• Beate Ritz, University of California, Los Angeles
• Luke Oman, NASA Goddard Space Flight Center
• Meredith Franklin, University of Southern California
• Theresa Bastain, University of Southern California
• Shohreh F. Farzan, University of Southern California
• Carrie Breton, University of Southern California
• Rima Habre, University of Southern California

Abstract

Accurate prediction of atmospheric air pollutants is critical for public health protection and environmental management. Traditional machine learning (ML) methods achieve high spatial resolution but lack physicochemical constraints, leading to systematic biases that compromise exposure estimates for epidemiological studies. Chemical transport models incorporate atmospheric physics but require expensive parameterization and often fail to capture local-scale variability crucial for health impact assessment. This gap between data-driven accuracy and physical realism presents a major obstacle to advancing air quality science. We address this challenge through a novel physics-informed deep learning framework that integrates advection–diffusion equations and fluid dynamics constraints directly into neural network architectures for multi-pollutant prediction. Our approach models air pollutant pairs across geographically distinct domains (NO₂/NO_x for California; PM_2.5/PM₁₀ for mainland China), providing a comprehensive framework for physics-constrained atmospheric modeling at high resolution. Through an efficient framework, our methodology demonstrates that incorporating proxy advection and diffusion fields as physical constraints fundamentally alters learning dynamics, reducing generalization error and eliminating systematic bias inherent in data-driven approaches while improving computational efficiency compared to graph networks. Site-based validation reveals unprecedented bias reduction: 21%–42% for nitrogen oxides and 16%–17% for particulate matter compared to the baseline deep learning methods. Our methodology uniquely generates physically interpretable parameters while providing explicit uncertainty quantification through ensemble techniques. The substantial bias reduction coupled with physically interpretable parameters has immediate implications for improving air pollutant exposure assessment and understanding in epidemiological research, potentially transforming health effect evaluations that rely on accurate spatial predictions.

Document Type

Article

Publication Date

12-1-2025

Notes/Citation Information

Publisher Copyright: © 2025 The Authors.

Digital Object Identifier (DOI)

10.1016/j.envint.2025.109915

Archival?

Archival

Repository Citation

Li, Lianfa; Khalili, Roxana; Lurmann, Frederick; Pavlovic, Nathan; Wu, Jun; Xu, Yan; Liu, Yisi; O’Sharkey, Karl; Ritz, Beate; Oman, Luke; Franklin, Meredith; Bastain, Theresa; Farzan, Shohreh F.; Breton, Carrie; and Habre, Rima, "Knowledge-informed deep learning to mitigate bias in joint air pollutant prediction" (2025). Epidemiology and Environmental Health Faculty Publications. 133.
https://uknowledge.uky.edu/epidemiology_facpub/133