The effort of properly accounting for multiplicity among stars has delivered many studies and thousands of catalogued systems, most recently dwarfed by about 800 thousand characterised binary stars by the ESA’s Gaia mission data release 3. This number, however, is 3 magnitudes lower than there are sources in the Gaia catalogue. This, coupled with the estimate that around 50% of all stars are actually a part of multi-star systems, indicates that far more binary stars are yet to be discovered according to the observed spectra.
Identifying binary stars through single-epoch photometry
This project is an exploratory study that aims to advance the identification of binary stars. It aims to detect binary stars through single-epoch photometry, whose precision is sufficient to constrain the spectral energy distribution (SED) of a given source at the level necessary (~mmag) for discriminating between single and binary stellar systems. It’s based on the idea that a SED of a binary stellar system has a unique shape that cannot be mistaken for a single-star SED. This might occur due to a temperature difference between the stars in a binary system, which will, provided the difference is large enough, result in a double-peak SED that should in principle be distinguishable from the single-peak SED.
Distinguishing between stellar systems on this basis enables the detection of otherwise hidden binary stars and the use of less expensive and readily available single-epoch photometric observations.
The development of Artificial Intelligence (AI) facilitated the adoption of the underlying machine learning (ML) techniques in many areas of Astrophysics. Specifically, it paved the way for faster and improved detection and analysis techniques in spectroscopy. In this project, we will use SED curves, derived from observed photometry, and 3D all-sky extinction maps coupled with parallax (distance from earth) measurements to train ML models to predict the probability of a source being either a single star or a binary system of stars.
Results
Our dataset contains photometric measurements of 10M stars, with 0.5% of them labelled as binary systems by other (non single-epoch photometry) surveys of the sky such as Gaia, Galah, Gaia-ESO, and Apogee. We experimented with several types of models, like gradient boosted decision trees, random forests and neural networks, as well as the addition of synthetic data in our training dataset. We compared the performance of our models on a subset of the real dataset. Naturally, models trained on the remainder of the real dataset performed best, since training and testing data were drawn from the same underlying distribution. During training, the difference between binary stellar systems and single stars is encoded in the model parameters. This enables the models to provide researchers with suggestions of possible binary systems.
The results are of course not ideal. Some binary systems are inherently difficult to classify. Specifically, binary systems with very similar stars, or systems, where one star is much brighter than the other, are indistinguishable from single stars. Therefore, we can only expect to identify binary systems that are sufficiently different, but not too different. Both gradient boosted decision trees and neural networks performed very similarly on all metrics. Prediction statistics are summarized in the confusion matrix.
Truth \ Predicted | binary | non-binary |
---|---|---|
binary | 0.10% | 0.35% |
non-binary | 0.21% | 99.34% |
About 21% of binary systems were correctly identified by the model, which could be somewhat improved by decreasing the probability threshold for classifying a stellar system as binary. Since the total fraction of stars in binary systems is estimated to be around 50% and not 0.5% as in our dataset, the accuracy with which those 0.5% are predicted is not of utmost importance and we believe the trained models will serve as a valuable tool for identifying potential binary stars for further study.
Partners
- Faculty of mathematics and Physics, University of Ljubljana
- European Space Agency (ESA)
Similar projects
- Photo2Chem - Stellar chemistry from photometric narrow-band imaging