Multiscale Methods for Data Assimilation in Turbulent Systems

摘要: Data assimilation of turbulent signals is an important challenging problem because the extremely complicated large dimension and incomplete partial noisy observations which usually mix scale mean flow small fluctuations. Due to limited computing power in foreseeable future, it desirable use multiscale forecast models are cheap fast mitigate curse dimensionality systems; thus model errors from imperfect unavoidable development a data method turbulence. Here we propose suite methods stochastic Superparameterization as model. seamless for parameterizing effect scales by local problems embedded coarse grid. The key ingredient systematic conditional Gaussian mixtures make efficient filtering subspace whose smaller than full state. proposed here tested on six dimensional conceptual dynamical turbulence mimics interesting features anisotropic including two way coupling between parts, intermittencies, extreme events Numerical results show that suitable have high skill estimating most energetic modes even with infrequent observation times.