Data Assimilation Research Team Events / Media

Machine learning has been receiving much attention to predict hydrometeorological disasters. Although many applications of machine learning focused on the fully data-driven and model-free prediction, the integration of a process-based model and machine learning has a potential to substantially improve the prediction of hydrometeorological disasters. In numerical weather prediction, machine learning has been applied to replace a part of an atmospheric model, such as a radiative transfer model, with a computationally cheaper data-driven model. In this talk, I will demonstrate that this surrogate modeling can contribute to reducing not only the computational cost but also the bias of process-based models and improving the prediction accuracy. I have two topics to present: (1) parameter optimization and uncertain quantification of a land surface model using surrogate modeling and Markov Chain Monte Carlo ( https://arxiv.org/abs/1909.04196) and (2) examining the potential of the surrogate of data assimilation’s analysis timeseries by ensemble Kalman filter and reservoir computing (https://arxiv.org/abs/2006.14276).