Date of Award:

12-2025

Document Type:

Thesis

Degree Name:

Master of Science (MS)

Department:

Computer Science

Committee Chair(s)

Xiaojun Qi

Committee

Xiaojun Qi

Committee

Hamid Karimi

Committee

Yang Shi

Abstract

Major solar flares are sudden, intense bursts of X-ray energy from the Sun, capable of severely impacting critical technological infrastructure like satellites, communication networks, and power grids on Earth. Accurate prediction of these high-intensity events is crucial but presents a significant challenge. This is largely due to their infrequent occurrence and the complex, dynamic nature of the Sun's underlying magnetic activity which drives these events. This research focuses on improving the prediction of major solar flares by utilizing detailed historical data that tracks the evolution of magnetic properties within solar active regions over time. This time-based data, however, contains inherent difficulties for building effective prediction models. Notably, major flares are significantly rarer than smaller events, the datasets often contain incomplete information or gaps, and the method used to collect the data can create complexities and redundancies that hinder accurate learning. We developed a new framework that combines a comprehensive, multi-step process to prepare this complex data with a computer model designed to overcome the dataset's challenges. It includes methods for handling missing information, making measurements comparable, focusing the model on clear examples by filtering less distinct data points, and importantly, generating additional examples of the rare major flare events to improve the model's ability to identify them. This process also incorporates careful data selection to account for how the data was collected over time. The computer model then learns to recognize the subtle magnetic precursors within this prepared data that indicate a future major flare. We rigorously tested our method using data from a period the model had not previously encountered. Evaluating its performance using a reliable measure of prediction skill particularly suited for rare events, our method achieved a high score exceeding 0.86. This result demonstrates a significant improvement compared to several other established prediction approaches. Our work underscores the critical importance of thorough data preparation and thoughtful design of learning systems in achieving robust and accurate forecasts for high-impact, low-frequency events in complex scientific domains such as space weather.

Checksum

71e4a9f1b88a55467ae195a5ab51c5f3

Share

COinS