Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Probabilistic neural network approach to determining parameters of eclipsing binaries

Marina Kounkel, Logan Sizemore, Hidemi Mitani Shen, Nicholas Chandler, Noah Reneau, Ian Pourlotfali, Ronald L. Payton, Brian Hutchinson, Ilija Medan, Keivan Stassun

Abstract

Eclipsing binaries provide one of the most direct mechanisms for measuring stellar properties such as mass and radius, but historically, determining these properties has been non-trivial and computationally prohibitive. As such, only a small fraction of all eclipsing binaries for which data have been available have been fully characterized. To improve computational efficiency, we construct an uncertainty-aware neural network which can ingest phase-folded light curves in any of 50 commonly used passbands, combined with phase-folded radial velocity measurements for both primary and secondary, as well as fluxes across the spectral energy distribution to predict stellar and orbital parameters of eclipsing binaries. The model was trained to be agnostic to the presence of third light, spots (both cool and hot), and incomplete data. As the model is operating in a probabilistic framework, it is also capable of outputting uncertainties in all of the parameters. The model was trained on synthetic data, and applied to a set of $\sim$200 previously solved real eclipsing binaries to demonstrate its performance. The model is capable of determining masses and radii of eclipsing binaries with precision of $\lesssim$20\% and $T_{\rm eff}$ with precision of $\sim$500 K in only a fraction of the time it takes the more traditional solvers. Although the resulting uncertainties are larger than what is possible to produce using more boutique analysis of individual stars, in the era of large photometric surveys, this approach allows to identify the most interesting systems, and it provides a starting point of the distributions in all of the parameters that these solvers could improve upon.

Probabilistic neural network approach to determining parameters of eclipsing binaries

Abstract

Eclipsing binaries provide one of the most direct mechanisms for measuring stellar properties such as mass and radius, but historically, determining these properties has been non-trivial and computationally prohibitive. As such, only a small fraction of all eclipsing binaries for which data have been available have been fully characterized. To improve computational efficiency, we construct an uncertainty-aware neural network which can ingest phase-folded light curves in any of 50 commonly used passbands, combined with phase-folded radial velocity measurements for both primary and secondary, as well as fluxes across the spectral energy distribution to predict stellar and orbital parameters of eclipsing binaries. The model was trained to be agnostic to the presence of third light, spots (both cool and hot), and incomplete data. As the model is operating in a probabilistic framework, it is also capable of outputting uncertainties in all of the parameters. The model was trained on synthetic data, and applied to a set of 200 previously solved real eclipsing binaries to demonstrate its performance. The model is capable of determining masses and radii of eclipsing binaries with precision of 20\% and with precision of 500 K in only a fraction of the time it takes the more traditional solvers. Although the resulting uncertainties are larger than what is possible to produce using more boutique analysis of individual stars, in the era of large photometric surveys, this approach allows to identify the most interesting systems, and it provides a starting point of the distributions in all of the parameters that these solvers could improve upon.

Paper Structure

This paper contains 11 sections, 7 figures.

Table of Contents

  1. Introduction
  2. Data
  3. Synthetic data
  4. Real data
  5. Model
  6. Model 1 (Tensorflow)
  7. Model 2 (PyTorch)
  8. Results
  9. Uncertainties
  10. Comparison between TensorFlow and PyTorch models
  11. Conclusions

Figures (7)

  • Figure 1: Distribution of the parameter space covered by the synthetic EBs.
  • Figure 2: Top: typical noise distribution applied to the synthetic data. Bottom: typical size of the dataset that was not masked in training.
  • Figure 3: Model diagrams showing the architectures of both models. The architecture of the models, as well as the hyperparameters associated with it were a result of various sweeps done to improve the overall performance.
  • Figure 4: Performance of the TensorFlow model on the real eclipsing binaries, showing a comparison between labels and the predictions for various parameters. Parameters of the primary star or the system as a whole are shown in blue, parameters of the secondary star are shown in red. The error bars show the uncertainties predicted by the model. The value in the top left corner of each panel shows the magnitude of the typical uncertainties in the sample. The inset histogram shows the characteristic scatter between the labels and the predictions, and the numbers in the top right corner indicate the mean and the standard deviation for the scatter.
  • Figure 5: Same as Figure \ref{['fig:tensorflow']}, but showing the performance of the PyTorch model.
  • ...and 2 more figures