Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EBOP MAVEN: A machine learning model to estimate the input parameters for analytic fitting of detached eclipsing binary light curves

Stephen Overall, John Southworth

TL;DR

ebop maven addresses the computational bottleneck in characterising detached eclipsing binaries by predicting JKTEBOP inputs from phase-folded light curves using a 1-D CNN with MC Dropout for uncertainties. Trained on a large synthetic dataset and validated on 20k synthetic cases and real DEBCat/TESS systems, the model achieves mean relative errors around $0.14$ on broad synthetic data, improving to $\sim$0.086 for deep eclipses and $8.7\pm0.7\%$ on real data, with $27$ of 28 real-system fits aligning with control analyses. The approach accelerates the generation of informed starting parameters for formal light-curve fitting, enabling scalable pipelines and catalogues for future missions like PLATO. Overall, ebop maven demonstrates feasibility of ML-assisted initial parameter estimation to streamline high-precision binary-star modelling in large time-domain surveys.

Abstract

Detached eclipsing binary stars (dEBs) are a key source of data on fundamental stellar parameters. Within the light curve databases of survey missions such as Kepler and TESS are a wealth of new systems awaiting characterisation. We aim to improve the scalability of efforts to process these data by developing a Convolutional Neural Network (CNN) machine learning model to assist in the automation of their analysis. From a phase-folded and binned dEB light curve the model predicts system parameters relating to stellar fractional radii, orbital inclination and eccentricity, and the stellar brightness ratio, for use as input values in subsequent formal analysis with the established JKTEBOP analytic code. We find the model able to predict these parameters for a previously unseen test dataset of 20000 synthetic dEB systems with a mean error of 14.1% when compared with the label values, improving to 8.6% against a subset representative of real systems. When tested with the TESS light curves of a set of real well-characterised systems, the model's predictions yield a mean error of $8.7\pm0.7\%$ when compared with label values derived from existing published analyses. Subsequent fitting of the TESS light curves with the JKTEBOP analytic code while using the model predictions as input values finds 27 of the 28 systems achieving a good fit. On the strength of these results, we plan to build a new characterisation pipeline based on the machine learning model and JKTEBOP code with the intention of producing a target catalogue of dEB systems for potential observation with the forthcoming PLATO mission.

EBOP MAVEN: A machine learning model to estimate the input parameters for analytic fitting of detached eclipsing binary light curves

TL;DR

ebop maven addresses the computational bottleneck in characterising detached eclipsing binaries by predicting JKTEBOP inputs from phase-folded light curves using a 1-D CNN with MC Dropout for uncertainties. Trained on a large synthetic dataset and validated on 20k synthetic cases and real DEBCat/TESS systems, the model achieves mean relative errors around on broad synthetic data, improving to 0.086 for deep eclipses and on real data, with of 28 real-system fits aligning with control analyses. The approach accelerates the generation of informed starting parameters for formal light-curve fitting, enabling scalable pipelines and catalogues for future missions like PLATO. Overall, ebop maven demonstrates feasibility of ML-assisted initial parameter estimation to streamline high-precision binary-star modelling in large time-domain surveys.

Abstract

Detached eclipsing binary stars (dEBs) are a key source of data on fundamental stellar parameters. Within the light curve databases of survey missions such as Kepler and TESS are a wealth of new systems awaiting characterisation. We aim to improve the scalability of efforts to process these data by developing a Convolutional Neural Network (CNN) machine learning model to assist in the automation of their analysis. From a phase-folded and binned dEB light curve the model predicts system parameters relating to stellar fractional radii, orbital inclination and eccentricity, and the stellar brightness ratio, for use as input values in subsequent formal analysis with the established JKTEBOP analytic code. We find the model able to predict these parameters for a previously unseen test dataset of 20000 synthetic dEB systems with a mean error of 14.1% when compared with the label values, improving to 8.6% against a subset representative of real systems. When tested with the TESS light curves of a set of real well-characterised systems, the model's predictions yield a mean error of when compared with label values derived from existing published analyses. Subsequent fitting of the TESS light curves with the JKTEBOP analytic code while using the model predictions as input values finds 27 of the 28 systems achieving a good fit. On the strength of these results, we plan to build a new characterisation pipeline based on the machine learning model and JKTEBOP code with the intention of producing a target catalogue of dEB systems for potential observation with the forthcoming PLATO mission.
Paper Structure (20 sections, 15 equations, 10 figures, 10 tables)

This paper contains 20 sections, 15 equations, 10 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Parameter selection
  4. Training and testing data
  5. Synthetic training dataset
  6. Synthetic testing dataset
  7. The test dataset of real systems
  8. The machine learning model
  9. Neural Networks
  10. Convolutional neural networks
  11. Model development and hyperparameter tuning
  12. Training and optimization
  13. Results
  14. Predictions on the synthetic test dataset
  15. Reviewing the scatter plot
  16. ...and 5 more sections

Figures (10)

  • Figure 1: Histograms showing the distribution of label values within the 1 million instances the full synthetic training dataset.
  • Figure 2: Histograms showing the distributions of label values within the 20 000 instances of the synthetic testing dataset.
  • Figure 3: The members of the test dataset of real systems plotted on a Hertzsprung-Russell diagram with a zero-age main-sequence (ZAMS) line derived from MIST 1.2 evolutionary models Dotter16apjsChoiDotter+16apj.
  • Figure 4: A fully connected neural network with three input neurons, two hidden layers and two output neurons.
  • Figure 5: Structure of the ebop maven model with the input to the left and outputs to the right. The input phase-folded light curve data is processed by six consecutive convolution blocks, each block consisting of a pair of 1-d convolutional layers (with the number of filters is shown) followed by a pooling layer which reduces the size of its output by a factor of 4. The output of the convolution blocks is flattened before being interpreted by a neural network made up of three dense layers interleaved with two dropout layers. Visualisation based on the PlotNeuralNet code Iqbal88plotneuralnet.
  • ...and 5 more figures