SIParCS 2026 - Guan Xin

A Novel Data Assimilation Technique to Better Resolve Ensemble Overconfidence

Recorded Talk

Ensemble numerical weather modelling coupled with Data Assimilation (DA) is a cornerstone of operational Numerical Weather Prediction (NWP). However, limited computational resources have forced Ensemble NWP to use computationally efficient but approximate procedures, leading to both model and sampling errors. In DA, a common technique employed to mitigate these errors is inflation, where the spread of an ensemble is increased to reduce model over-confidence. Common inflation algorithms have difficulty fully mitigating abovementioned errors, while some also have difficulty adjusting to rapidly changing observing systems. In this work, I present a novel inflation algorithm that aims to better mitigate model and sampling errors by incorporating inflation as part of the DA analysis scheme. This also intrinsically provides an in-built adaptability to the spatiotemporal heterogeneity of observations. Experiments were carried out in the Data Assimilation Research Testbed (DART), on the Lorenz-96 low-order model over a range of experimental conditions. The proposed novel algorithm is shown to consistently out-perform the standard DART adaptive inflation algorithm (Anderson 2009, El Gharamti 2018) in terms of RMSE, and frequently in terms of ensemble consistency. These results validate the value of testing this algorithm in more realistic weather models, which if successful, could provide a means to improve operational weather prediction.

Mentors: Moha El Gharamti, Jeffrey Anderson

Slides and poster