Development of a Global Climate Model for Atmospheric Temperature Using Machine Learning

Abstract

This study presents a novel three-dimensional global model of atmospheric temperature developed using Artificial Neural Networks (ANNs) trained on radio occultation (RO) data from the COSMIC I and COSMIC II satellite missions. Over 14.7 million quality-controlled profiles were used, providing approximately 9.5 billion data points that capture temperature variability across latitude, longitude, altitude (0-60 km), and time (2006-2025). The global domain was divided into 1,296 spatial grid cells (10° × 5°) to enable localized ANN training and ensure efficient handling of regional atmospheric dynamics. Model performance was evaluated through cross-validation and independent testing against radiosonde measurements from 684 stations worldwide. Results show mean absolute errors (MAE) of 1.5-4.5°C and root-mean-square errors (RMSE) of 2.5-6.5°C, with best performance in the tropical troposphere and increasing errors toward high latitudes. The model successfully reproduces key climatological structures (including the tropospheric lapse rate, stratospheric inversion, and seasonal hemispheric asymmetries), and accurately captures diurnal and annual thermal cycles. Long-term simulations (2006-2025) reveal a distinct tropospheric warming trend (∼ +0.07°C per year at 11 km) and a corresponding stratospheric cooling (∼ -0.03°C per year near 32 km), consistent with established satellite and reanalysis records. These results demonstrate that ANN-based frameworks can effectively model global atmospheric thermal structure and evolution, providing a scalable approach for future climate monitoring and forecasting applications.