This study examines the minimum ensemble size for accurate geophysical forecasting using a method called the ensemble Kalman filter. We reformulate accuracy via observation noise-dependency to classify filter performance qualitatively. Through numerical experiments with a chaotic model, we link the minimum ensemble size for the accuracy to system's instability and propose an effective ensemble downsizing method that ensures both stability and accuracy.

Recent wildfire activity in Canada has been rising the public demand for fast, yet accurate operational air quality forecasts along with related forecast uncertainties. This study explores the potential of an efficient process-based approach to estimate forecast uncertainties of operational air quality models. Our results show its potential for understanding how uncertainties of operational forecast evolve over time and an improved use of sparse observation signals at remote wildfire regions.

The recharge oscillator model has served as a simple yet powerful toy model for modelling the El Niño–Southern Oscillation. Whereas the addition of Gaussian noise is sufficient to capture the observed overall statistical features, it fails to resolve dynamical signatures associated with major El Niño events. Here we propose to employ a conditional non-Gaussian noise model that better captures the effect of large westerly wind bursts and their effect on major warming events.

The risk of climate change triggering tipping points is a concern. While Earth System Models (ESMs) predict these points, their timing often varies. Equations of nonlinear dynamical systems can model tipping, hysteresis, and inertia, enabling comparison with ESMs. We propose a method to map tipping timing estimates onto a dynamical system that simulates a full hysteresis loop. We focus on solutions to the equation under overshoot forcing in global temperature and different inertia levels.

Deep learning has been used to discover new data assimilation methods to track chaotic dynamical systems. Strikingly, these data assimilation networks can match the accuracy of ensemble-based methods using only a single state forecast. This paper first investigates the reasons for this efficacy, and then shows that their performance in more nonlinear regimes matches that of the most accurate scalable data assimilation methods, with greater efficiency and robustness.

The reliability of Bayesian calibration depends on the quality and availability of observational data. But are we choosing the right data? We address this question by measuring the information gained during calibration to quantify how data selection influences the Bayesian calibration of physics-based landslide runout models. We find that more data does not always yield better results – observations that capture the dynamics governed by a parameter are more effective for its calibration.

Recent increases in the variability of Antarctic sea ice have elicited much interest and research on these changes. Here, we examine observations taken from three specific years (2016, 2021, and 2023) which either contain or precede a period of anomalously low sea ice extent. To understand the combined influence of weather systems, surface temperatures, and atmospheric pressure on sea ice formation and decay, we apply novel methods from machine learning and dynamical systems.

We developed a method to plan tiny, local nudges that reduce spikes in weather-like simulations. It was motivated by the lack of reliable ways to design small interventions in unpredictable systems. The method uses many parallel forecasts to suggest where a nudge matters, then tests and refines it with repeated full-model simulations. In two models, it reduced extremes and peak regional wind speeds with similar or smaller input, enabling safer study of intervention ideas.

The goal of data assimilation in geoscience is to merge information from models and observations. In the presence of structures such as ocean eddies, classic assimilation schemes fail to merge those two sources, resulting in position and shape errors. Here, we propose a new scheme which assimilates only the parameters describing a simplified eddy structure. Comparison with classic schemes shows that our approach preserves the eddy parameters and may reduce the computational cost.