Title: Imputing Missing Long-Term Spatiotemporal Multivariate Atmospheric Data with CNN-Transformer Machine Learning Show Chinese title

Title: 使用CNN-Transformer机器学习填补缺失的长期时空多变量大气数据

Abstract: Continuous physical domains are important for scientific investigations of dynamical processes in the atmosphere. However, missing data arising from operational constraints and adverse environmental conditions pose significant challenges to accurate analysis and modeling. To address this limitation, we propose a novel hybrid Convolutional Neural Network (CNN) Transformer machine learning model for multivariable atmospheric data imputation, termed CT-MVP. This framework integrates CNNs for local feature extraction with transformers for capturing long-range dependencies across time and altitude. The model is trained and evaluated on a testbed using the Specified Dynamics Whole Atmosphere Community Climate Model with thermosphere and ionosphere extension (SD-WACCM-X) dataset spanning 13 years, which provides continuous global coverage of atmospheric variables, including temperature and zonal and meridional winds. This setup ensures that the ML approach can be rigorously assessed under diverse data-gap conditions. The hybrid framework enables effective reconstruction of missing values in high-dimensional atmospheric datasets, with comparative evaluations against traditional methods and a simple transformer. The results demonstrate that CT-MVP achieves superior performance compared with traditional approaches, particularly in cases involving extended periods of missing data, and slightly outperforms a simple transformer with the same hyper-parameters. Show Chinese abstract

Abstract: 连续的物理领域对于大气中动态过程的科学研究非常重要。 然而，由于操作限制和恶劣环境条件导致的数据缺失，给准确的分析和建模带来了重大挑战。 为解决这一限制，我们提出了一种新型的混合卷积神经网络（CNN）Transformer机器学习模型，用于多变量大气数据填补，称为CT-MVP。 该框架将CNN用于局部特征提取，同时利用Transformer捕捉时间与高度上的长程依赖关系。 该模型在使用指定动力学全大气社区气候模型与热层和电离层扩展（SD-WACCM-X）数据集的测试平台上进行训练和评估，该数据集覆盖13年，提供了包括温度和纬向及经向风在内的大气变量的连续全球覆盖。 这种设置确保了在多种数据缺口条件下对ML方法进行严格评估。 该混合框架能够有效重建高维大气数据集中的缺失值，并与传统方法和简单的Transformer进行了比较评估。 结果表明，与传统方法相比，CT-MVP表现出更优的性能，特别是在长时间数据缺失的情况下，且在相同超参数下略优于简单的Transformer。