Mechanism-Constrained Learning for Regime-Aware Urban Ozone Prediction and Interpretation

Accurately characterizing source-related and precursor factors associated with surface ozone (O3) pollution remains a central challenge for air quality management. Data-driven models can achieve strong predictive performance, but mechanistically consistent interpretation remains challenging. Here we develop a mechanism-constrained diagnostic framework (MCDF) integrating observations, source-related factors of volatile organic compounds (VOCs) resolved by positive matrix factorization, and scenario-specific constraints from the empirical kinetic modeling approach (EKMA). Its core mechanism-constrained multitask model (MCMM) predicts O3 while jointly learning precursor-control regime and signed ridge distance, with a direction-consistency loss guiding precursor responses toward the corresponding photochemical state. MCMM achieved stable repeated-split O3 prediction performance and produced finite precursor responses more consistent with EKMA than an unconstrained model. Across 10-30% increases in nitrogen oxides (NOx), MCMM predicted negative O3 responses in 90.1-92.3% of initially VOCs-limited cases. Applied to observations from a representative O3-polluted megacity case in China, MCDF revealed regime-dependent changes in model-attributed precursor and source-related associations. Overall, MCDF integrates scenario-specific photochemical constraints with data-driven O3 prediction, bridges predictive performance and mechanistic interpretation, and provides an interpretable, regime-aware framework for diagnosing precursor sensitivities and source-related variability in complex urban O3 pollution.