Oral

IAMAS

JMP02 - Machine Learning in atmospheric, ocean and earth-system prediction: forecasting, simulation and scientific analysis

Data-driven global atmosphere-ocean-land coupled model

Deep learning-driven data models have recently gained significant attention in global weather forecasting within 2 weeks. However, the current approaches introduced so far have limitations to extend its forecasts beyond two weeks due to their inability to accurately simulate atmosphere-ocean-land interactions. This study develops a deep learning-based atmosphere-ocean-land coupled model to improve long-range climate simulation/forecast accuracy. To properly simulate coupling strength between the atmosphere, ocean, and land modules, the one module's encoders to predict its own components are separately trained from the encoders to be coupled to other modules (so called 'coupled-feature generation (CG) module'). After the pre-training of each sub-coupled modules (i.e., atmosphere + ocean CG modules, ocean + atmosphere CG modules, etc.), those are fine-tuned within a fully coupled model framework. In addition, the coupled model is designed to predict tendency (i.e., difference between the future and current values) to less rely on the autoregressive feature of predictands and replay buffer mechanism is applied to increase model's ability to predict long-term climate. The proposed model is compared to a data-driven coupled model in which all atmosphere-ocean-land variables are trained and predicted as a single model. Inference for a century exhibited stable climate states without any significant drift, confirming its ability to simulate long-term climate. The simulation quality of El Nino-Southern Oscillation, which is one of most prominent atmosphere-ocean coupled process, is compared with the observations. The proposed deep learning coupled model exhibited higher performance for predicting the heatwaves in regions with strong atmosphere-land coupling.