Physics‐Guided CNN‐LSTM Model With Multi‐Head Attention for Aerosol Optical Depth Prediction

Abstract Accurate aerosol optical depth (AOD) prediction remains challenging due to complex aerosol‐radiation interactions and highly variable spatio‐temporal patterns. Three critical scientific issues motivate this work: understanding whether and how physical principles can enhance deep learning predictions, identifying which aerosol properties most strongly govern AOD variations, and improving the prediction of extreme AOD events critical for air quality management. Herein, utilizing MERRA‐2 reanalysis data (1980–2024) over the Huaihe River Basin in eastern China, a Physics‐Guided deep learning framework is presented for Aerosol Optical Depth (AOD) prediction. The model proposed integrates Convolutional Neural Networks (CNN), Long Short‐TermMemory (LSTM) networks, and multi‐head attention mechanisms to capture both spatio‐temporal features and physical relationships of aerosol properties. Three key aspects are involved: First, a hybrid deep learning model is developed and evaluated, which combines CNNs for spatial correlation extraction, bidirectional LSTM for temporal dependency modeling, and multi‐head attention for feature interaction learning. Second, a comprehensive feature importance analysis is conducted by examining the relationships between different aerosol properties (mass concentration, scattering coefficient, and Ångström exponent) and AOD prediction, offering physical insights into the model's decision‐making process. Third, a specialized approach is proposed for extreme AOD event prediction, focusing on early detection and accurate forecasting of high‐AOD episodes. Overall, the results demonstrate the model's efficacy in capturing both regular AOD variations and extreme events, with the Physics‐Guided architecture showing superior performance compared to traditional methods. This integrated approach enhances AOD prediction accuracy and deepens insights into aerosol‐radiation interactions, thereby improving atmospheric monitoring and air quality forecasting. While MERRA‐2 has inherent temporal delays, this framework provides valuable capabilities for historical trend analysis, numerical model validation, and can be readily adapted for real‐time applications through transfer learning with satellite observations.