AtmosChemNet-AI for chemistry-informed urban air pollution forecasting using spatiotemporal deep learning

Urban air pollution prognosis is an important public health protection and environmental management challenge that continues to endure due to the complex, nonlinear interactions among emissions, meteorology, transport, and atmospheric chemistry. Recent deep learning techniques have been designed to directly improve short-term predictive skill by learning spatiotemporal correlations from historical data. However, these data-driven, often implicit approaches essentially bypass chemical processes. Consequently, they tend to perform poorly under regime shifts, i.e., high-pollution episodes, photochemically active conditions, and unobserved meteorological patterns, thereby limiting their robustness and interpretability. In this paper, we propose AtmosChemNet-AI, a reaction-kinetics-aware deep learning framework, to circumvent these limitations and enable the atmospheric science community to predict the dynamics of urban atmospheric pollution. Our new method incorporates chemistry-inspired reaction-kinetics-aware proxy descriptors into learning via kinetics-informed feature construction, kinetics-conditioned gating in a spatiotemporal sequence model, and physics-constrained regularisation during training. Coupled, these mechanisms constrain representation learning to chemically feasible pollutant evolution while maintaining the expressiveness of modern deep neural networks. Validation is performed across a range of pollutants, forecast horizons, and operating regimes using real data, urban air quality observations, and meteorological reanalysis. Experimental evaluation under the considered datasets, forecasting horizons, and baseline comparisons demonstrates consistent improvements in predictive performance, including 15–20% relative reductions in mean absolute error and root mean square error, together with higher coefficients of determination. Ablation and regime analyses corroborate the finding that knowledge of reaction kinetics increases robustness during high-pollution events, under reaction-favourable conditions, and in untrained meteorological regimes. The main results show that incorporating reaction-kinetics domain knowledge into deep learning models notably increases forecast accuracy, stability, and generalisation. This framework provides a promising foundation for future real-time air quality forecasting and urban environmental decision-support applications, subject to further deployment-oriented validation and system-level evaluation.