[M-GI26] Data-driven approaches for weather and hydrological predictions

[MGI26-06] Earth System Modeling in Latent Space: A Time-Varying Parameter Approach

*澤田 洋平^1,2、- Le Duc^1,2 (1.東京大学、2.気象庁気象研究所)

キーワード：潜在空間、データ同化、パラメータ最適化、状態空間モデル

Data assimilation, an essential technique in Earth system prediction such as weather forecasting, can be recognized as a state-space model which integrates observation into process-driven models to sequentially estimate latent variables, or state variables in the models. In data assimilation, the dynamics of these latent variables and their relationship with observations are prescribed by process-driven models such as atmospheric models, enabling efficient estimation of latent variables with limited observations. However, the rise of fully data-driven approaches, particularly those employing deep neural networks, recently marks a shift in Earth system modeling. Unlike data assimilation, these data-driven models require learning the dynamics of latent variables and the mapping from observations onto latent variables directly from data, posing challenges in contexts with sparse observation. Consequently, many data-driven models use data assimilation outputs as "observations." When Earth system prediction is recognized as modeling and estimating latent variables, functions of parameters in physically based process-driven models have yet to be extensively discussed. In our preliminary work, we propose a novel approach that combines process-driven and data-driven modeling, utilizing time-varying model parameters. We recognize physically explainable time-varying model parameters (e.g., parameters in convection triggers of a cumulus parameterization) as meaningful latent variables. Then, we suggest the efficient estimation of dynamics of these time-varying parameters by the extension of conventional ensemble data assimilation. In this presentation, we aim to interpret current process-driven and data-driven prediction methods from the perspective of latent space modeling and invite discussions on the future direction of Earth system modeling, emphasizing the integration of process understanding and data-driven insights.