Probabilistic neural network approach to determining parameters of eclipsing binaries

This paper presents a probabilistic neural network that rapidly and efficiently estimates stellar masses, radii, and effective temperatures for eclipsing binaries from light curves and radial velocity data with moderate precision, offering a scalable solution for characterizing large survey datasets and prioritizing systems for detailed follow-up analysis.

The Gist

The Big Picture: The "Black Box" Problem

Imagine you are a detective trying to solve a mystery: What are two stars made of, how big are they, and how heavy are they?

The stars are locked in a cosmic dance, orbiting each other so closely that they occasionally eclipse (block) one another from our view. By watching how the light dims and wobbles, astronomers can figure out the stars' secrets.

The Problem: Doing this mathematically is like trying to solve a Rubik's Cube while blindfolded, on a shaking boat, and with a timer that counts down in days. Traditional methods require running complex simulations (called MCMC) that can take days or weeks on supercomputers for just one pair of stars. With millions of stars being discovered by new telescopes, we are drowning in data but starving for answers.

The Solution: The authors built a Neural Network (a type of AI) called EBNet. Think of this AI as a "Cosmic Speed Reader." Instead of solving the math from scratch every time, it has "read" millions of practice problems and learned to recognize patterns instantly. It can tell you the mass, size, and temperature of a star pair in less than one second.

---

How the AI Was Trained: The "Video Game" Analogy

You can't teach an AI by showing it real stars immediately because real stars are messy, and we don't always know the "right answer" for them.

So, the team built a simulated universe (using a program called PHOEBE).

• The Training Data: They generated 300,000 fake star systems.
• The Chaos: They didn't make these fake stars perfect. They added "noise" (static), "spots" (like sunspots on the fake stars), and "third light" (a sneaky background star peeking in).
• The Goal: They taught the AI to look at the messy, fake light curves and say, "Ah, I know this pattern! That's a star with a 50% chance of being a giant and a 20% chance of being a dwarf."

By training on this chaotic "video game" world, the AI learned to ignore the noise and focus on the real signal, just like a seasoned detective ignores the red herrings in a mystery novel.

---

How It Works: The "Swiss Army Knife" Approach

Most old tools were like specialized screwdrivers: one tool for light, another for speed. If you didn't have the right tool, the job got stuck.

EBNet is a Swiss Army Knife. It can eat almost anything you throw at it:

1. Light Curves: The graph of brightness over time (from any color filter, like red, blue, or infrared).
2. Radial Velocities: How fast the stars are wobbling toward or away from us.
3. Spectral Energy Distribution (SED): The total "rainbow" of light the star emits.

The Magic Trick:
If you give it a full dataset (light + speed + color), it gives a very precise answer.
If you give it only the light curve (no speed data), it doesn't crash. Instead, it says, "I can still guess, but I'm not 100% sure, so here is a wider range of possibilities."

• Analogy: Imagine guessing someone's height. If you see them standing next to a door (full data), you know they are 6 feet tall. If you only see their shadow on a wall (partial data), you might guess "between 5 and 7 feet." EBNet is smart enough to tell you how wide that guess is.

---

The Results: Fast, Good Enough, and Honest

The team tested their AI on about 200 real star systems that humans had already solved the hard way.

• Speed: It was instant. What took humans days took the AI milliseconds.
• Accuracy: It got the mass and radius within about 20% of the true value. For effective temperature, it was off by about 500 degrees.
  • Is 20% error bad? In the world of astronomy, where we are dealing with objects light-years away, this is actually a fantastic starting point. It's like getting a weather forecast that says "It will rain" instead of "It will rain at 2:03 PM." It's good enough to know if you need an umbrella.
• The "Uncertainty" Superpower: This is the most important part. The AI doesn't just guess a number; it guesses a range of confidence. If the data is messy, the AI admits, "I'm not sure." If the data is clean, it says, "I'm pretty confident."

Why This Matters: The "Filter" for the Future

We are entering an era where telescopes like TESS, LSST, and ZTF will find millions of these binary stars. We cannot possibly hire enough human astronomers to run the slow, expensive "days-long" math on every single one.

EBNet acts as a filter:

1. It scans millions of stars in seconds.
2. It flags the "boring" ones (where the data is too messy or the stars are uninteresting).
3. It highlights the "gems"—the systems that look weird, interesting, or perfect for study.
4. It gives human astronomers a great starting guess so that when they do run the slow, expensive math, they can do it in minutes instead of days.

The Bottom Line

The authors built a fast, probabilistic AI that acts as a "first responder" for binary stars. It's not perfect, and it's not a replacement for the detailed, boutique analysis of a single star. But in a universe overflowing with data, it is the tool that allows us to stop drowning and start swimming, helping us find the most interesting stars to study in the first place.

Technical Summary

1. Problem Statement

Double-lined spectroscopic eclipsing binaries (EBs) are critical astrophysical laboratories for directly measuring stellar mass, radius, and effective temperature ($T_{\text{eff}}$). However, determining these parameters is computationally expensive and non-trivial.

• Current Limitations: Traditional methods (e.g., PHOEBE) rely on Markov Chain Monte Carlo (MCMC) sampling to find optimal fits and uncertainties. This process can take days per system, even on compute clusters, and requires reasonable initial guesses.
• Existing ML Tools: Previous neural network estimators like EBAI (2008) are fast but suffer from significant limitations:
  • They are restricted to specific passbands (V-band), causing bias in modern surveys (TESS, Kepler, Gaia) with custom filters.
  • They are deterministic, offering no uncertainty estimates.
  • They are decoupled from radial velocity (RV) data, preventing simultaneous optimization of light curves (LCs) and RVs.
  • They struggle with complex system features like starspots, third light contamination, and incomplete data.

2. Methodology

The authors developed EBNet, a probabilistic neural network designed to ingest phase-folded light curves, radial velocities, and spectral energy distribution (SED) fluxes to predict stellar and orbital parameters with associated uncertainties.

A. Data Generation (Training Set)

Since real solved EBs are insufficient for training a robust model, the authors generated $\sim300,000$ synthetic systems using PHOEBE:

• Parameter Space: Randomly drawn distributions for $T_{\text{eff}}$ (2500–12,000 K), mass ratio ($q$), period (log-normal), radii, inclination, eccentricity, and semi-major axis.
• Complexity Injection: The training data intentionally included:
  • Spots: Cool and hot spots on primary and secondary stars.
  • Third Light: Contamination from background or tertiary stars.
  • Missing Data: Random masking of RV measurements, light curves, and SED fluxes to simulate real-world incompleteness.
  • Noise: Gaussian noise profiles applied to simulate varying signal-to-noise ratios across different instruments.
• Inputs: 50 passbands (Kepler, TESS, Gaia, etc.), 512 phase-binned points per light curve, RV curves for both components, and 31 SED bandpasses.

B. Model Architecture

Two parallel models were developed and trained on the same data but with different frameworks and hyperparameters:

1. Model 1 (TensorFlow): Uses aggressive convolution and pooling to reduce 50 light curves into a single channel representation, followed by fully connected layers with residual connections.
2. Model 2 (PyTorch): Utilizes a deeper stack of 2D convolutional blocks with Group Normalization and ReLU activations to process Light Curves, RVs, and SEDs separately before concatenation.

Probabilistic Output: Both models output 42 values: 21 parameters (mass, radius, $T_{\text{eff}}$, $q$, $e$, $\omega$, etc.) and their corresponding 21 variances. They are trained using a Gaussian Negative Log-Likelihood (NLL) loss function, allowing the network to learn the distribution of parameters rather than a single deterministic point.

C. Ensemble Strategy

Since neither model outperformed the other across all parameters (PyTorch was better at $T_{\text{eff}}$; TensorFlow was more robust to noise for other parameters), the authors created a hybrid system. The TensorFlow weights were converted to PyTorch (via ONNX), and the final EBNet evaluates both models, selecting the best prediction for each specific parameter.

3. Key Contributions

• Probabilistic Framework: Unlike previous estimators, EBNet provides uncertainty estimates for all parameters, reflecting the intrinsic noise and data quality.
• Multi-Modal Input: The model simultaneously processes light curves from 50 different passbands, radial velocities, and SED fluxes, making it agnostic to the specific survey or filter used.
• Robustness to Imperfections: The model is explicitly trained to handle spots, third light, and missing data types (e.g., LCs without RVs).
• Speed: Predictions are generated in <1 second per system, compared to days for MCMC-based solvers.

4. Results

The model was tested on a catalog of 220 real, previously solved eclipsing binaries.

• Accuracy:
  • Mass and Radius: Achieved precision of $\lesssim 20\%$.
  • Effective Temperature ($T_{\text{eff}}$): Achieved precision of $\sim 500$ K (PyTorch model).
  • Correlation: Predictions strongly follow the 1-1 trend line with literature values.
• Uncertainty Calibration:
  • The predicted uncertainties generally align with the scatter between predictions and ground truth (typical scatter is $0.7\sigma$ to $0.85\sigma$).
  • While the model's reported uncertainties ($\sim 20\%$) are larger than boutique literature values ($<2\%$), this is attributed to the model's generalist nature and the inclusion of systematic errors often ignored in individual studies.
• Data Scenarios:
  • Without RVs: The model can still estimate radii and orbital ratios, though mass ratio ($q$) and eccentricity ($e$) become less certain. The model correctly increases uncertainty estimates in these cases.
  • Without SED: $T_{\text{eff}}$ precision degrades significantly ($\sim 5000$ K uncertainty), confirming the necessity of SED data for temperature constraints.
  • Third Light/Spots: The model successfully recovers parameters despite these contaminants, which are often difficult for traditional solvers to disentangle without good initial guesses.

5. Significance and Future Impact

• Survey Scalability: EBNet is uniquely suited for the "data deluge" from large photometric surveys (Kepler, TESS, ZTF, LSST). It can rapidly characterize thousands of systems that would be impossible to solve individually with MCMC.
• Initial Guess Generator: The model provides high-quality initial guesses for traditional solvers (like PHOEBE), significantly reducing the time required for MCMC convergence.
• System Selection: By providing rapid, uncertainty-aware estimates, EBNet allows astronomers to efficiently identify the most scientifically interesting systems for follow-up spectroscopy or detailed boutique analysis.
• Distribution Calibration: It enables the creation of statistical distributions of binary parameters across large populations, serving as a baseline for evolutionary studies.

In conclusion, EBNet represents a paradigm shift from slow, manual, or boutique fitting of eclipsing binaries to a fast, automated, and uncertainty-aware probabilistic approach, bridging the gap between massive survey data and precise stellar astrophysics.