Hybrid Physics-Informed Machine Learning Framework for Interpretable Ozone Formation Modelling in a Coastal Urban Location

Abstract Surface ozone (O₃) formation is governed by highly nonlinear interactions among precursor emissions, meteorological conditions, and photochemical processes, making accurate prediction and interpretation challenging. This study presents a hybrid data-driven framework that integrates machine learning (XGBoost and Long Short-Term Memory networks), Physics-Informed Neural Networks (PINNs), and Symbolic Regression (SR) to model ozone dynamics over a coastal urban environment in Kannur in south India, using high-resolution (15-minute) observations. Among the models, XGBoost achieved the highest predictive accuracy (R² = 0.98; RMSE = 0.60 ppbv), effectively capturing nonlinear relationships between ozone and its precursors. The data-driven PINN demonstrated comparable accuracy (R² = 0.98; RMSE = 0.89 ppbv) while ensuring physical consistency by embedding photochemical constraints within the learning process. Symbolic regression further provided interpretable analytical expressions that revealed key ozone production and loss pathways, consistent with established atmospheric chemistry, including NOₓ titration, photolysis, and aerosol-mediated suppression. Feature attribution using SHAP identified solar radiation, temperature, and NO₂ as dominant drivers of ozone formation, while particulate matter exhibited a moderating effect through radiative attenuation. The results highlight that hybrid physics-guided machine learning approaches can simultaneously achieve high predictive accuracy, physical realism, and chemical interpretability. This study demonstrates that integrating data-driven models with physical constraints and equation discovery offers a robust and scalable framework for atmospheric chemistry modeling, with significant implications for air quality forecasting, environmental policy, and climate–air quality interactions.