Date of Award

2026

Document Type

Thesis

Degree Name

Bachelor of Science

Department

Physics

Abstract

Leveraging the unprecedented data volumes from modern astronomical surveys, this work addresses the critical challenge of identifying catastrophic outliers (COs) in photometric redshift (photo-z) estimates. Reliable redshift estimation is essential for precision cosmology, yet COs (galaxies with severely erroneous photo-z) can introduce significant systematic bias. We develop and evaluate two machine learning approaches for robust CO identification: a multilayer perceptron (MLP) utilizing photo-z point estimates and probability distributions, and a convolutional neural network (CNN) leveraging raw 5-band photometric images. Using datasets emulating LSST conditions, we demonstrate that the MLP model achieves high accuracy and stability across redshift bins, correctly identifying up to 97.5% of COs while misclassifying less than 1% of non-COs. The CNN model, though still under development, shows promising potential to bypass traditional preprocessing and directly extract relevant features from imaging data. We report that the CNN model correctly identified up to 82.3% of COs while misidentifying 4.1% of incorrectly identified NOs. We implement and assess techniques such as data rebalancing, bin weighting, and CO weighting to address sample imbalances and optimize model performance. Our results confirm and extend previous findings, highlighting the value of photo-z probability distributions and advanced neural architectures for outlier mitigation.

These methods are broadly applicable to future large-scale surveys, paving the way

for more automated and reliable photo-z quality control in cosmological analyses.

Download

Included in

Physics Commons

Share

COinS