Files and links (1)
PDFOpen Access (License Unspecified), Open Access
Dissertation
DeepVoid: physics-aware semantic segmentation of large-scale structure with deep learning
Doctor of Philosophy (Ph.D.), Drexel University
Sep 2025
DOI:
https://doi.org/10.17918/00011185
Abstract
Cosmic voids, the dominant volume-filling structures of the universe, provide sensitive probes of dark energy, modified gravity, and the growth of structure. Their identification, however, depends strongly on the chosen algorithm, tracer population, and smoothing scale, leading to inconsistent catalogs across methods. This dissertation presents DeepVoid, a physics-informed deep learning framework for semantic segmentation of the cosmic web in three dimensions. Using voxel-wise tidal tensor classifications as ground truth, we train 3D U-nets to label each voxel as void, wall, filament, or halo directly from the mass density field. We first demonstrate that U-nets trained on dark matter density fields can reproduce physically defined voids with high fidelity, even under sparse tracer conditions with intertracer separations of order ~ 10 h⁻¹ Mpc. Building on this foundation, we extend DeepVoid with refinements aimed at bridging the gap to observational data. These include training on redshift-space distorted inputs, testing alternative loss functions to mitigate class imbalance, and adding astrophysical inputs such as galaxy color and flux. We explore conditioning strategies based on tracer separation, incorporate attention modules into the U-Net, and develop a systematic progressive curriculum learning scheme with multiple validation strategies. Model interpretability is examined through activation and attention maps, and robustness is tested through boundary analyses and cross-simulation validation. Together, these developments show that deep neural networks can approximate physically motivated void definitions while remaining flexible to survey-like conditions, providing a foundation for the use of voids as cosmological probes in forthcoming galaxy surveys.
Metrics
8 File views/ downloads
6 Record Views
Details
- Title
- DeepVoid
- Creators
- Samuel F. Kumagai
- Contributors
- Michael Scott Vogeley (Advisor)
- Awarding Institution
- Drexel University
- Degree Awarded
- Doctor of Philosophy (Ph.D.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xii, 108 pages
- Resource Type
- Dissertation
- Language
- English
- Academic Unit
- College of Arts and Sciences; Physics; Drexel University
- Other Identifier
- 991022093052104721