Issue |
A&A
Volume 691, November 2024
|
|
---|---|---|
Article Number | A106 | |
Number of page(s) | 17 | |
Section | Catalogs and data | |
DOI | https://doi.org/10.1051/0004-6361/202449995 | |
Published online | 01 November 2024 |
Application of Convolutional Neural Networks to time domain astrophysics. 2D image analysis of OGLE light curves
1
Departamento de Astronomía, Universidad de La Serena,
Avenida Juan Cisternas 1200,
La Serena,
Chile
2
European Southern Observatory,
Karl-Schwarzschild-Str. 2,
85748
Garching,
Germany
3
Gemini Observatory, NSF’s NOIRLab,
Av. J. Cisternas 1500 N,
1720236
La Serena,
Chile
4
Cerro Tololo Inter-American Observatory/NSF’s NOIRLab,
Casilla 603,
La Serena,
Chile
★ Corresponding author; nicolas.monsalves@userena.cl
Received:
15
March
2024
Accepted:
9
August
2024
In recent years the amount of publicly available astronomical data has increased exponentially, with a remarkable example being large-scale multiepoch photometric surveys. This wealth of data poses challenges to the classical methodologies commonly employed in the study of variable objects. As a response, deep learning techniques are increasingly being explored to effectively classify, analyze, and interpret these large datasets. In this paper we use two-dimensional histograms to represent Optical Gravitational Lensing Experiment phasefolded light curves as images. We use a Convolutional Neural Network (CNN) to classify variable objects within eight different categories (from now on labels): Classical Cepheid, RR Lyrae, Long Period Variable, Miras, Ellipsoidal Binary, Delta Scuti, Eclipsing Binary, and spurious class with Incorrect Periods (Rndm). We set up different training sets to train the same CNN architecture in order to characterize the impact of the training. The training sets were built from the same source of labels but different filters and balancing techniques were applied. Namely: Undersampling, Data Augmentation, and Batch Balancing (BB). The best performance was achieved with the BB approach and a training sample size of ~370 000 stars. Regarding computational performance, the image representation production rate is of ~76 images per core per second, and the time to predict is ~60 μs per star. The accuracy of the classification improves from ~92%, when based only on the CNN, to ~98% when the results of the CNN are combined with the period and amplitude features in a two step approach. This methodology achieves comparable results with previous studies but with two main advantages: the identification of miscalculated periods and the improvement in computational time cost.
Key words: methods: data analysis / methods: statistical / catalogs / surveys / time / stars: variables: general
© The Authors 2024
Open Access article, published by EDP Sciences, under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
This article is published in open access under the Subscribe to Open model. Subscribe to A&A to support open access publication.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.