Physics-Regularized Operator Learning for Hourly Mapping of Near-Surface Air Pollution under Sparse Observations

Abstract

Hourly air-pollution fields are needed for exposure assessment and air-quality management; however, ground monitors are sparse, and satellite observations are intermittent. Data-driven models may reproduce station observations while leaving unmonitored regions poorly constrained. This mismatch raises a central concern for observation-free hourly mapping: station-level skill may not imply a physically reliable pollution field. To test this, we trained a neural operator to predict near-surface CO over Mainland China from emissions, meteorology, and terrain, using ground observations only for training. Transport regularization was imposed through a full-grid advection-diffusion-reaction residual that balances emissions, transport, diffusion, and decay. On the unseen year 2022, this constraint increased hourly station agreement from R = 0.57 to 0.64 and structural agreement with an assimilation-informed reanalysis from R = 0.55 to 0.74. In a blind-region test, where local monitors were removed from training, R increased from 0.51 to 0.58. Transport regularization therefore reduces nonphysical extrapolation and enables more stable hourly pollution-field reconstruction without observations at prediction time.