1: \begin{abstract}
2: We propose a high-order stochastic-statistical moment closure model
3: for efficient ensemble prediction of leading-order statistical moments
4: and probability density functions in multiscale complex turbulent
5: systems. The statistical moment equations are closed by a precise
6: calibration of the high-order feedbacks using ensemble solutions of
7: the consistent stochastic equations, suitable for modeling complex
8: phenomena including non-Gaussian statistics and extreme events. To
9: address challenges associated with closely coupled spatio-temporal
10: scales in turbulent states and expensive large ensemble simulation
11: for high-dimensional systems, we introduce efficient computational
12: strategies using the random batch method (RBM). This approach significantly
13: reduces the required ensemble size while accurately capturing essential
14: high-order structures. Only a small batch of small-scale fluctuation
15: modes is used for each time update of the samples, and exact convergence
16: to the full model statistics is ensured through frequent resampling
17: of the batches during time evolution. Furthermore, we develop a reduced-order
18: model to handle systems with really high dimension by linking the
19: large number of small-scale fluctuation modes to ensemble samples
20: of dominant leading modes. The effectiveness of the proposed models
21: is validated by numerical experiments on the one-layer and two-layer
22: Lorenz '96 systems, which exhibit representative chaotic features
23: and various statistical regimes. The full and reduced-order RBM models
24: demonstrate uniformly high skill in capturing the time evolution of
25: crucial leading-order statistics, non-Gaussian probability
26: distributions, while achieving significantly lower computational cost
27: compared to direct Monte-Carlo approaches. The models provide effective
28: tools for a wide range of real-world applications in prediction, uncertainty
29: quantification, and data assimilation.
30: \end{abstract}
31: