Date of Award

Spring 2026

Language

English

Embargo Period

4-21-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

College/School/Department

Department of Atmospheric and Environmental Sciences

Program

Atmospheric Science

First Advisor

Andrea Lopez Lang

Second Advisor

Zheng Wu

Committee Members

Ryan Torn, Paul Roundy

Keywords

forecasting, subseasonal, temperatures, extremes, Northern Hemisphere

Subject Categories

Applied Statistics | Atmospheric Sciences | Longitudinal Data Analysis and Time Series | Meteorology | Multivariate Analysis | Non-linear Dynamics | Other Statistics and Probability

Abstract

Wintertime stratospheric dynamics provide key information for understanding atmospheric teleconnections and improving subseasonal-to-seasonal (S2S) predictions on timescales of two weeks to two months. Periods of enhanced predictability, often referred to as forecasts of opportunity, arise from large-scale teleconnected variability, within which the stratosphere serves as an important precursor for tropospheric states, such as near-surface temperatures. While traditional diagnostics of downward coupled stratosphere-troposphere interactions typically rely on zonal-mean representations of wind and geopotential height, this dissertation presents an alternative vortex-centric framework through metrics that capture the daily geometric and dynamical evolution of the stratospheric polar vortex. The proposed stratospheric polar vortex ellipse metric approach identifies a best-fit ellipse defining the polar vortex’s edge and approximates its shape and position. These metrics provide a physically consistent and computationally efficient method for quantifying vortex variability and offer a novel set of predictors for diagnosing the downward influence of the stratosphere.

The core of this dissertation evaluates the utility of the stratospheric polar vortex ellipse metrics for identifying S2S forecasts of opportunity across the Northern Hemisphere. Using ERA-5 reanalysis for winter seasons (November to March) from 1959/1960 to 2021/2022, machine learning (ML) models of varying complexity are compared to establish a reproducible framework for predicting the sign of 1000-hPa temperature anomalies at a 14-day lead time. Results demonstrate that Random Forest (RF) models leveraging the proposed ellipse metrics produce skillful and physically interpretable predictions, performing with probabilistic accuracy comparable to more complex ML architectures (e.g., Long Short-Term Memory networks). An eXplainable AI technique, Shapley Additive Explanations (SHAP), was used to further understand the RF predictions, revealing that input features that the RF models used for high-confidence predictions align with the dynamical understanding.

The regional dependence of these forecasts for opportunity on S2S timescales is further explored by extending the RF model to three distinct regions: Northeast Europe, Northeast Canada, and the Southeast United States, with longer lead times. While probabilistic forecast accuracy generally decreases with lead time, the features that were important to the predictions vary by region. Analysis of SHAP values and predictor distributions during high-confidence predictions confirms that distinct manifestations of stratospheric variability drive these regional differences in S2S forecast skill.

Finally, this dissertation evaluates the representation of the stratospheric polar vortex within a state-of-the-art S2S reforecast system: NOAA’s Unified Forecast System (UFS) coupled model prototypes. Given that these UFS prototypes vary in vertical extent and resolution and physical parameterizations, such as gravity wave representations, this assessment highlights how the model design impacts stratospheric forecast utility. The results show that most UFS prototypes successfully capture key characteristics of stratospheric variability and its downward coupling. Furthermore, incorporating UFS-derived vortex metrics into the RF model further improves UFS near-surface temperature predictions on subseasonal timescales, highlighting the potential of hybrid dynamical and ML approaches to enhance S2S forecast models.

Ultimately, this dissertation establishes an interpretable framework for leveraging simple metrics of stratospheric variability in ML models to identify key stratospheric features that are associated with confident predictions of tropospheric-level temperatures and extremes. This collection of results may benefit improved subseasonal outlooks and encourage the development of a more resolved stratosphere in future coupled Earth systems models. This work demonstrates the fidelity and utility of applying a relatively simple ML model to identify and evaluate forecasts of opportunity from the stratosphere for surface S2S prediction.

License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download

Share

COinS