Advances in air quality modeling through artificial intelligence, machine learning, and deep learning: A...

Air quality modeling is undergoing a significant transformation, as machine learning (ML) and deep learning (DL) emerge to address the inherent limitations of traditional methods, including statistical approaches and chemical transport models (CTMs). This review provides a systematic overview of this evolution, categorizing advances from 112 publications into two main approaches: purely data-driven models for estimating pollutant concentrations, and ML-assisted numerical models that refine CTM outputs or emulate the models themselves. Collectively, the literature reveals that data-driven approaches are significantly improving forecast accuracy for key pollutants like fine particulate matter (PM_2.5) and ozone. Concurrently, ML-assisted methods are enhancing traditional modeling. Bias correction efforts improve model accuracy for more reliable downstream applications like source apportionment, while DL-based emulators offer high-fidelity simulations with substantially reduced computational cost. Despite such progress, the application of ML to air quality modeling still faces significant challenges, such as the "black-box" nature of models, inadequate uncertainty quantification, and regional data scarcity. Furthermore, integrating the stiff chemical ordinary differential equations that govern atmospheric physics remains a critical hurdle. Addressing these issues requires emerging solutions. eXplainable AI (XAI) enhances model transparency, while Physics-Informed Neural Networks (PINN) enable the fusion of data-driven inference with fundamental physical laws, guiding the development of faster, more reliable, and interpretable air quality modeling systems. Ultimately, these advancements are driving the development of next-generation air quality models capable of providing more timely and accurate information for assessing pollution impacts and guiding mitigation strategies.