Solving 150,000+ Eclipsing Binary Light Curves With An AI-Powered Model

NASA's Kepler and Tess missions provided us with unprecedented photometric data, both in terms of precision and in terms of temporal coverage. Of particular interest are light curves of eclipsing binary stars (EBs) as their modeling allows us to determine their physical and orbital properties from geometry and well-understood laws of dynamics. The main bottleneck of their analysis is the computation of the forward-model, i.e. synthesizing a theoretical light curve from a set of parameters. In this proposal we demonstrate that a neural network, trained on synthetic data computed with a state-of-the- art EB modeling code, can provide near-identical results while delivering a whopping 7+ orders of magnitude speed-up. This effectively reduces the time required for sampling of parameter posteriors from days to seconds, allowing us to rigorously and robustly model hundreds of thousands of observed EB light curves in a comparatively short time. In particular, the network is trained on physical and geometric parameters as input and syntesized light curves as output; when fully trained, the network "synthesizes" light curves in milliseconds, irrespective of the parametric complexity associated with a particular model. Thus, instead of running a computationally expensive model for millions of times as part of the sampling process, we run the AI-model for millions of times instead, achieving the aforementioned speed-up. We retain the analytical rigor of obtaining parameter posteriors and marginalizing over nuisance parameters with minimal impact to accuracy. The proposal aims to: (1) construct an extensive, general training dataset based on observed parameter distributions; (2) optimize neural network topology (i.e., the number of layers, the number of units per layer, the choice of the activation function) to maximize light curve similarity; (3) estimate photometric parameters for 150,000+ EB light curves from Kepler and from TESS, including those extracted from full-frame images; (4) perform quality control of derived parameters and assess any dependence of parameter distributions on the training set; (5) extend the training set to cover any identified parameter regions with inadequate coverage; (6) re-estimate photometric parameters for the entire sample of 150,000+ EBs; and (7) analyze parameter distributions and ingest them into the catalog on the dedicated webpage and on MAST as a High-Level Science Product. The default parameter set will be augmented for the cases where EB morphology allows it (for example, photometric mass ratios can be derived for contact binaries), where we have auxiliary parameters (for example, Gaia parallaxes allow us to determine absolute stellar luminosities), or auxiliary datasets (for example, RVs for double-lined EBs allow us to derive all absolute fundamental parameters for both components). The resulting catalog promises to be the most extensive, objectively sampled set of EBs available to date, and the developed methodology will be applicable to other ongoing and upcoming missions such as Gaia or Plato. The project will provide training to 7 junior researchers and future investigators.