Akila successfully defended on Monday, March 9, 2026
Throughout her time with us, Akila's dedicated contributions have made her an invaluable member of both the iHARP and UMBC communities. This significant achievement is a true testament to her scholarly rigor, hard work, and immense talent.
Congratulations, Akila, on this incredible milestone! We have been honored to witness your growth and look forward to your future successes. The entire iHARP community wishes you the very best as you embark on your next great adventure!
- Dr. Jianwu Wang, Chair/Advisor (UMBC)
- Dr. Vandana Janeja, Co-Chair (UMBC)
- Dr. Houbing Song (UMBC)
- Dr. James Foulds (UMBC)
- Dr. Donald K Perovich (Dartmouth College)
- Dr. Nicole Schlegel (NOAA)
Rapid and accelerating Arctic climate change poses significant challenges for artificial intelligence (AI) systems, primarily due to severe data scarcity, inherent nonlinearities, and complex spatiotemporal interactions within the ocean–ice–atmosphere system. Conventional deep learning approaches rely heavily on large data volumes and often lack physical consistency. Consequently, purely data-driven models may produce physically implausible predictions and offer limited interpretability, thereby reducing their utility for scientific discovery and decision-making. To enable reliable and trustworthy sea ice prediction, physics-embedded learning architectures are required to leverage domain-specific priors that encode known physical laws, constraints, and causal mechanisms.
This dissertation presents a physics-integrated deep learning framework for modeling and analyzing the time-series evolution of Arctic sea ice. The proposed framework systematically combines physical knowledge with data-driven learning to enhance predictive performance, interpretability, and scientific validity. Specifically, this work introduces three complementary strategies for embedding physical constraints and governing principles directly into the learning process, each addressing a distinct scientific challenge in Arctic climate research.
First, a physics-informed deep learning model is developed for sea ice thickness prediction by explicitly incorporating thermodynamic constraints and governing energy-balance laws into the loss function. This approach ensures that model predictions respect known physical relationships while remaining flexible enough to learn from sparse observational data. Second, the dissertation introduces physics-encoded neural network architectures that embed established physical relationships directly into the model structure. These architectures enable the inference of latent physical parameters from noisy, incomplete proxy data, facilitating physically meaningful representation learning. Third, knowledge-guided temporal causal models are formulated to quantify the causal impact of sea ice variability on coupled oceanic and atmospheric processes. By incorporating time-varying treatments, causal structural constraints, and physics-based priors, these models provide interpretable estimates of causal effects that are consistent with established physical mechanisms rather than spurious correlations.
Across all evaluation settings, the results demonstrate that physics-integrated models consistently outperform conventional deep learning baselines. Overall, this work highlights the critical role of physics-integrated machine learning in advancing predictive capability, causal understanding, and trustworthiness in climate and Earth system modeling.