Data Mining for Gravitationally Lensed Quasars

TL;DR

This paper presents a machine learning-based method for efficiently identifying gravitationally lensed quasars in large survey data, significantly improving purity and completeness over traditional techniques.

Contribution

It introduces a two-step machine learning approach combining catalog-level neural networks and pixel-level pattern recognition, enhancing candidate selection for future large-scale surveys.

Findings

Achieves 70% purity and 60% completeness after candidate selection.

Reduces candidate list processing time to a few CPU hours.

Outperforms simple colour cut methods in purity and efficiency.

Abstract

Gravitationally lensed (GL) quasars are brighter than their unlensed counterparts and produce images with distinctive morphological signatures. Past searches and target selection algorithms, in particular the Sloan Quasar Lens Search (SQLS), have relied on basic morphological criteria, which were applied to samples of bright, spectroscopically confirmed quasars. The SQLS techniques are not sufficient for searching into new surveys (e.g. DES, PS1, LSST), because spectroscopic information is not readily available and the large data volume requires higher purity in target/candidate selection. We carry out a systematic exploration of machine learning techniques and demonstrate that a two step strategy can be highly effective. In the first step we use catalog-level information ($griz$+WISE magnitudes, second moments) to preselect targets, using artificial neural networks. The accepted targets are then inspected with pixel-by-pixel pattern recognition algorithms (Gradient-Boosted Trees), to form a final set of candidates. The results from this procedure can be used to further refine the simpler SQLS algorithms, with a twofold (or threefold) gain in purity and the same (or $80\%$) completeness at target-selection stage, or a purity of $70\%$ and a completeness of $60\%$ after the candidate-selection step. Simpler photometric searches in $griz$+WISE based on colour cuts would provide samples with $7\%$ purity or less. Our technique is extremely fast, as a list of candidates can be obtained from a stage III experiment (e.g. DES catalog/database) in {a few} CPU hours. The techniques are easily extendable to Stage IV experiments like LSST with the addition of time domain information.