This work extends the Mapping Particle Filter to account for local dependencies. Two localization methods are tested: a global particle flow with local kernels, and iterative local mappings based on correlation radius. Using a two-scale Lorenz-96 truth and a one-scale forecast model, experiments with full/partial and linear/nonlinear observations show Root Mean Square Error reductions using localized Gaussian mixture priors, achieving competitive performance against Gaussian filters.

Artificial intelligence (AI) has recently shown promising results in ENSO (El Niño–Southern Oscillation) forecasting, outperforming traditional models. Yet AI models deliver accurate predictions without showing the underlying mechanisms. Our study examines a specific AI model, the reservoir computer (RC). Our results show that the RC is less sensitive to initial perturbations than the traditional Zebiak–Cane (ZC) model. This reduced sensitivity can explain the RC's superior skills.