Deep learning for air pollutant forecasting: opportunities, challenges, and future directions

Abstract Deep learning methods are increasingly employed to forecast air quality from an ever-increasing stream of data and algorithms. However, the efficacy of current approaches may be questionable when evaluated not solely in terms of greater forecasting fidelity, but also concerning the decision-making process in pollution early warning. Here, rather than amending classical machine learning algorithms, we argue that now is the time to push the frontiers of air pollutant forecasting beyond state-of-the-art approaches. This can be achieved through near real-time assimilation of multi-scale observations for laying the foundation of training data, enhanced attribution methods for impending heavy pollution, diagnostics for forecasting uncertainty, and advanced climate-chemistry emulators for improving seasonal forecasting. To harness this potential, it is essential to address several key challenges in deep learning methods, particularly generalization ability in extreme events, physics-informed interpretable approaches, and the mitigation technology of cumulative errors in multi-process coupled systems. This interdisciplinary endeavor will remain a central pursuit in the quest to anticipate and manage environmental change.