Date of Award

2025-12-01

Degree Name

Doctor of Philosophy

Department

Computer Science

Advisor(s)

Olac Fuentes

Abstract

Deep neural networks excel at a wide range of processing tasks across various disciplines. However, the quantity and quality of data significantly impact network performance. In specialized domains, high-quality datasets are often difficult to gather, interpret, and curate for effective learning. Few-shot learning techniques, including transfer learning, data augmentation, and meta-learning, have emerged to address these constraints.

We propose three physics-guided strategies for enhancing neural networks trained with limited data: (1) combining existing models like LSTMs with Physics-Informed Neural Networks through two-branch architectures that merge their outputs, (2) deriving custom physics-informed data augmentation algorithms to expand limited datasets, and (3) adding physics-informed terms to the loss functions of existing models like U-Nets. These strategies are applied to two challenging problems: Fluid Flow Velocity Prediction in mechanical engineering and Glacier Segmentation in geology.

For fluid flow prediction, we developed a novel Physics-Informed LSTM architecture that achieved 73.7% improvement in U-velocity and 85.3% improvement in V-velocity prediction compared to LSTM-only approaches. For glacier segmentation, our physics-informed data augmentation improved clean-ice IoU from 68.17 to 71.22 (a 4.5% gain) and debris-covered ice IoU from 35.94 to 45.92 (a 27.8% gain) over the baseline. We also developed a physics-informed velocity loss that fuses glacier velocity fields to promote physically consistent segmentations. The combination of all our glacier segmentation strategies achieved a state-of-the-art Debris-Covered Ice IoU of 46.07%, representing a 28.2% relative improvement over prior benchmarks. To enable our velocity physics experiments and open the door for future higher-resolution studies, we also built a production-ready, generalizable glacier-velocity dataset creation pipeline that fuses ITS_LIVE dynamics with Landsat/ICIMOD data.

This research demonstrates that integrating domain knowledge of physics into neural networks provides a viable pathway for improving performance on data-limited problems where well-understood physical laws with mathematical representations can guide the learning process.

Language

en

Provenance

Received from ProQuest

File Size

86 p.

File Format

application/pdf

Rights Holder

Jose Guadalupe Perez Zamora

Download

Share

COinS