Robust ensemble feature extraction for solving conditional nonlinear optimal perturbation

Conditional nonlinear optimal perturbation CNOP has been widely applied to predictability and sensitivity studies of nonlinear models. The popular methods of solving CNOP can be divided into two categories: adjoint-based and ensemble-based. Although the adjoint-based method is very accurate, it requires the development of adjoint models. The ensemble-based method is an adjoint-free technique, but either its robustness is weak or its filtering process is dependent on observation or experience to a great extent. In this paper, we propose a robust ensemble-based method to solve CNOP. This method does not need the filtering process and can extract robust features. To demonstrate the validity, the proposed method is applied to the Zebiak-Cane ZC model and compared with the adjoint-based method and other ensemble-based methods. To improve the computational efficiency, we design an OpenMP-based parallelising scheme for the proposed method. Experimental results show that the proposed method can outperform other ensemble-based methods in robustness and the corresponding solution of CNOP significantly approximates the one obtained with the adjoint-based method.