Parallel PCA-Based Bacterial Foraging Optimization Algorithm for Identifying Optimal Precursors of North Atlantic Oscillation

The North Atlantic Oscillation (NAO) as a prominent atmospherical mode is of great research value since its significant impact on global climate change. The contributing factor of NAO events is research hotspot of atmospheric sciences. The study of optimal precursor (OPR) which can trigger the NAO events is conducive to improving the predictability. Conditional nonlinear optimal perturbation (CNOP) method is a useful tool to solve this kind of nonlinear initial value problem, and has been widely applied in identifying OPRs of different climate events with various numerical models. In this work, the OPRs of NAO events are identified using Community Earth System Model (CESM). However, solving CNOP in such a high-dimension model with data scale of more than 5 millions is a relatively difficult and computationally expensive task. Thus, we extend the dimensionality reduction idea to a swarm intelligence algorithm called Bacterial Foraging Optimization Algorithm (BFOA) to explore the OPRs of NAO events. The task is transformed into the problem in low-dimensional space with principal component analysis (PCA). To speed up the process, we implement the parallelization of algorithm and atmospheric component. Message Passing Interface (MPI) is adopted to make bacteria search in parallel, and Compute Unified Device Architecture (CUDA) is utilized to port massive matrix computation onto GPUs. With NVIDIA Tesla K80 on the Tianhe-2 supercomputer, the PCA-based BFOA (PBFOA) achieves a speed-up of 41.8× compared with its serial version. Moreover, the OPRs obtained by the proposed PBFOA can trigger the NAO events stably. The results indicate that the proposed method can solve CNOP in complicated models effectively and fleetly.