The capability of deep learning model to predict atmospheric compositions across spatial and temporal domains

Machine learning (ML) techniques have been extensively applied in the field of atmospheric science. It provides an efficient way of integrating data and predicting atmospheric compositions. However, whether ML predictions can be extrapolated to different domains with significant spatial and temporal discrepancies is still unclear. Here we explore the answer to this question by presenting a comparative analysis of surface carbon monoxide (CO) and ozone (O₃) predictions by integrating deep learning (DL) and chemical transport model (CTM) methods. The DL model trained with surface CO observations in China in 2015-2018 exhibited good spatial and temporal extrapolation capabilities, i.e., good surface daily CO predictions in China in 2019-2020 and over 10% independent observation stations in China in 2015-2020. The spatial and temporal extrapolation capabilities of DL model are further evaluated by predicting hourly surface O₃ concentrations in China, the United States (US) and Europe in 2015-2022 with a DL model trained with surface O₃ observations in China and the US in 2015-2018. Compared to baseline O₃ simulations using GEOS-Chem (GC) model, our analysis exhibits mean biases of 2.6 and 4.8 µg/m³ with correlation coefficients of 0.94 and 0.93 (DL); and mean biases of 3.7 and 5.4 µg/m³ with correlation coefficients of 0.95 and 0.92 (GC) in Europe in 2015-2018 and 2019-2022, respectively. This analysis indicates the potential of DL to make reliable atmospheric composition predictions over spatial and temporal domains where a wealth of local observations for training is not available.