Nonlinear local Lyapunov vectors (NLLVs) are developed to indicate orthogonal directions in phase space with different perturbation growth rates. In particular, the first few NLLVs are considered to be an appropriate orthogonal basis for the fast-growing subspace. In this paper, the NLLV method is used to generate initial perturbations and implement ensemble forecasts in simple nonlinear models (the Lorenz63 and Lorenz96 models) to explore the validity of the NLLV method. The performance of the NLLV method is compared comprehensively and systematically with other methods such as the bred vector (BV) and the random perturbation (Monte Carlo) methods. In experiments using the Lorenz63 model, the leading NLLV (LNLLV) captured a more precise direction, and with a faster growth rate, than any individual bred vector. It may be the larger projection on fastest-growing analysis errors that causes the improved performance of the new method. Regarding the Lorenz96 model, two practical measures—namely the spread–skill relationship and the Brier score—were used to assess the reliability and resolution of these ensemble schemes. Overall, the ensemble spread of NLLVs is more consistent with the errors of the ensemble mean, which indicates the better performance of NLLVs in simulating the evolution of analysis errors. In addition, the NLLVs perform significantly better than the BVs in terms of reliability and the random perturbations in resolution.

The dropsondes released during the Tropical Cyclone Intensity (TCI) field campaign provide high-resolution kinematic and thermodynamic measurements of tropical cyclones within the upper-level outflow and inner core. This study investigates the impact of these upper-level TCI dropsondes on analyses and prediction of Hurricane Patricia (2015) during its rapid intensification (RI) phase using an ensemble–variational data assimilation system. In the baseline experiment (BASE), both kinematic and thermodynamic observations of TCI dropsondes at all levels except the upper levels are assimilated. The upper-level wind and thermodynamic observations are assimilated in additional experiments to investigate their respective impacts. Compared to BASE, assimilating TCI upper-level wind observations improves the accuracy of outflow analyses verified against independent atmospheric motion vector (AMV) observations. It also strengthens the tangential and radial wind near the upper-level eyewall. The inertial stability within the upper-level eyewall is enhanced, and the maximum outflow is more aligned toward the inner core. Additionally, the analyses including the upper-level thermodynamic observations produce a warmer and drier core at high levels. Assimilating both upper-level kinematic and thermodynamic observations also improves the RI forecast. Compared to BASE, assimilating the upper-level wind induces more upright and inward-located eyewall convection, resulting in more latent heat release closer to the warm core. This process leads to stronger inner-core warming. Additionally, the initial warmer upper-level inner core produced by assimilating TCI thermodynamic observations also intensifies the convection and latent heat release within the eyewall, thus further contributing to the improved intensity forecasts.

Quantifying the predictability limits of chaotic systems and their forecast models is an important issue with both theoretical and practical significance. This paper introduces three invariant statistical properties of attractors, namely the attractor radius, global attractor radius (GAR), and the global average distance between two attractors, to define the geometric characteristics and average behavior of a chaotic system and its error growth. The GAR is \(\sqrt 2\) times the attractor radius. These invariant quantities are applied to quantitatively measure the global and local predictability limits (both have practical and potential predictability limits, which correspond to the attractor radius and GAR, respectively) of both global ensemble average forecasts and one single initial state, respectively. Both the attractor radius and GAR are intrinsic properties of a chaotic system and independent of the forecast model and model errors, and thus provide more accurate, objective metrics to assess the global and local predictability limits of forecast models compared with the traditional error saturation or asymptotic value (AV). Both the Lorenz63 model and operational forecast data are used to demonstrate the theoretical aspects of these geometric characteristics and evaluate the feasibility and effectiveness of their application to predictability analysis.

Traditional ensemble forecasting based on numerical weather prediction (NWP) models, is constrained by the need for massive computational resources, resulting in limited ensemble sizes. Although emerging artificial intelligence (AI)-based weather models offer high forecast accuracy and improved computational efficiency, they still face considerable challenges in ensemble forecasting applications, due to the unclear error growth dynamic and the lack of suitable ensemble methods in AI-based models. In this study, we propose a fast, physics-constrained perturbation scheme through the self-evolution dynamics of an AI-based weather model for ensemble forecasting of tropical cyclones (TCs). These initial perturbations are conditioned on specific amplitude and spatial characteristics, exhibiting physically reasonable dynamical growth and spatial covariance. Based on this perturbation scheme, the TC track ensemble forecasts within the AI-based model significantly outperform those from the European Centre for Medium-Range Weather Forecasts (ECMWF) for both deterministic and probabilistic metrics. Notably, we conduct TC track forecasts with 2000 members for the first time, achieving further enhanced forecast skills in probability distribution and extreme scenarios of TC movement.