[M-GI27] Data-driven approaches for weather and hydrological predictions

[MGI27-11] Toward enhancing the ensemble Kalman filter with a diffusion model

*本田 匠¹、澤田 洋平² (1.東京大学情報基盤センター、2.東京大学大学院工学系研究科)

キーワード：データ同化、機械学習、数値天気予報、拡散モデル

Data assimilation provides improved initial conditions by combining information from observations and simulations taking their uncertainties into account. The ensemble Kalman filter (EnKF), a major data assimilation method, leverages differences within an ensemble of simulations to represent forecast uncertainty. To accurately estimate forecast uncertainty, a large ensemble is necessary, but this can be computationally expensive. Recent studies have indicated that diffusion models, a type of generative artificial intelligence (AI), can generate a large ensemble at relatively low computational cost. This study aims to integrate a diffusion model into an EnKF system to improve state estimation accuracy. Specifically, we use a diffusion model to generate a large ensemble using a small conventional ensemble simulation as input. By doing so, we mitigate sampling errors due to limited ensemble sizes. As a proof-of-concept, this study focuses on a low-dimensional chaotic system. We built a diffusion model based on a U-net neural network and trained it on a small conventional ensemble dataset. Preliminary results indicate that our diffusion model successfully generates a large ensemble that captures the general characteristics of the forecast error covariance. Next, we will evaluate the performance of an EnKF system coupled with the trained diffusion model.