Finding Galaxies Beyond: Vision Transformer Based Gravitational Lens Search Model Using Synthetic Data
CSEF · 2023 Physics & Astronomy Third Award
Overview
Gravitational lenses are rare but significant phenomena in astronomy, where a foreground galaxy distorts the light of a distant galaxy. Each lens enables the study of dark matter, the Hubble Constant, and early galaxies. Only a fraction of estimated lenses have been found. My goal is to build an image classification model to find lenses in telescope data. Unlike existing lens search models that often use Convolutional Neural Networks trained on either real or synthetic data on supercomputer GPU clusters, I use a Vision Transformer model trained on both real and simulated lenses and non-lenses in 40 minutes on one Nvidia Titan GPU. The model uses self-attention to find relationships between parts of the image. I predicted that this would be optimal for lens images which involve a lens and source shape. Uniquely, I simulated synthetic lens images of different configuration/distortion types, developing the different astrophysical parameter distributions. The training set consists of lens and non-lens images from Dark Energy Survey, Hyper-Suprime Cam, Sloan Digital Sky Survey(SDSS) in addition to simulated ones. Training with only real images resulted in an accuracy score of ~96% on a control test set. Trained with both real and synthetic data, the accuracy was ~98%. Finally, I applied the trained model on 60,000 images from the SDSS and identified 13 previously undiscovered lens candidates that I have confirmed with a JPL NASA researcher. I created a website for any researcher to use my model and find lenses in their image data.
Source coverage
This record comes from a published award list, not a complete project archive. Its abstract comes from CSEF's public project showcase as archived by the Internet Archive before judging (https://web.archive.org/web/20230401224130/https://ca-csef.zfairs.com/showcase/ShowcaseInfo?f=838e60b7-ea75-46e8-865c-fde4864244b3); the version presented may differ.
Awards (1)
Competition history
- CSEF 2023
Resources
Related projects
ISEF · 2023
Discovering Gravitationally Lensed Galaxies Using a Vision Transformer Image Classification Model Trained on Synthetic and Real Data
ISEF · 2026
Curriculum Learning-Based Detection and Analysis of Dark Matter via Strong Gravitational Lensing
ISEF · 2022
Analysis of Ring Galaxies Detected Using Deep Learning With Real and Simulated Data
CSEF · 2023
Discovering Rare Ca II Quasar Absorption Lines with a Neural Network
CWSF · 2026
Discovering Hidden Galaxy Mergers: Finding Mergers Missed by Detection Methods
ISEF · 2026
Discovering Hidden Galaxy Populations in TNG100-1 With Unsupervised Autoencoder Learning
ISEF · 2018
Challenging Limitations: Using Deep Learning, Time Series Analysis, and Statistical Methods for Noise Reduction to Develop an Innovative Approach to Exoplanet Candidate Detection Using Earth-Based Telescopes
ISEF · 2017
A Linked Learning Approach to Automated Galaxy Morphology Classification
Closest projects by meaning, across every fair and year in the corpus.
Browse more like this
Source: California Science & Engineering Fair public projects