A physics-informed multi-variable transformer model for imputing missing sodium LiDAR observations over the Andes

Light Detection and Ranging (LiDAR) has emerged as a key technology for obtaining high-resolution measurements of atmospheric conditions, particularly in the mesosphere-lower-thermosphere (MLT) region. However, gas in data, caused by environmental limitations and operational interruptions, pose significant challenges for analysis and modeling. To overcome this issues, we introduce PI-MVTI, a novel physics-informed, multi-variable transformer-based imputation model. This hybrid framework combines convolutional neural networks (CNNs) for extracting local features with transformers to model long-range temporal and vertical dependencies. Additionally, physical constraints based on continuity and buoyancy principles are embedded into the model to ensure that the imputed values are both accurate and physically consistent. Comparative results demonstrate that PI-MVTI outperforms traditional interpolation methods, especially in scenarios with extended periods of missing data.