Reconstructing particle number size distributions via aerosol history–informed multiple deep learning models
摘要
Particle number size distribution (PNSD) is fundamental for characterizing aerosols and quantifying aerosol–cloud interactions, while simulations remain uncertain due to strong spatiotemporal variability during air-parcel transport. We developed a framework that combines air-parcel location history with co-located aerosol, cloud, meteorological, and gas variables into deep learning (DL) models to reconstruct PNSD. Two Long Short-Term Memory (LSTM) models and one Bidirectional LSTM (BiLSTM) model were trained and evaluated on 10 years of measurements at the Cape Verde Atmospheric Observatory (CVAO) in the central Atlantic, achieving mean fractional errors (MFEs) < 0.17. Applying the three well-trained models to the Ascension Island (ASI) dataset in the South Atlantic achieved MFE < 0.16, demonstrating strong transferability and enabling PNSD projection in remote marine regions. SHapley Additive exPlanations (SHAP) analysis showed inconsistent feature attributions across three models, suggesting limited interpretability of mechanisms during air-parcel transportation. Thus, multi-model cross-validation is recommended for interpreting DL models.