Date of Award

2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Committee Chair

Tathagata Mukherjee

Committee Member

Letha Etzkorn

Committee Member

Jacob Hauenstein

Research Advisor

Tathagata Mukherjee

Subject(s)

Solar activity--Computer simulation, Deep learning (Machine learning), Artificial intelligence

Abstract

Foundation Models (FM) are large neural networks trained in a self-supervised fashion for a specific pretraining objective, to learn a general representation. FMs have become the default paradigm in many language and vision tasks extending to scientific domains, which have been seeing promising results. Yet in heliophysics, most models are still task specific and are constrained by scarce labeled data. In this work, we trained Surya, an FM with 366M parameters. We use the Solar Dynamics Observatory (SDO) observations, including eight Atmospheric Imaging Assembly (AIA) channels and five Helioseismic and Magnetic Imager (HMI) products at their full resolution of 4096 × 4096. Surya is based on spatiotemporal transformer architecture trained with a pretraining objective of forecasting on the full resolution, further optimized with an autoregressive rollout tuning. This pretraining strategy allowed Surya to perform zero shot forecasting of solar dynamics and flare events. We fine-tuned the model using a Low-Rank Adaption (LoRA) method demonstrating strong performance in downstream tasks.

Available for download on Thursday, August 05, 2027

Share

COinS