Astronomers have developed a new artificial intelligence system that can automatically identify a rare and unusual class of binary stars known as heartbeat stars. The method, which uses a recurrent neural network, successfully sifted through astronomical data to find these stellar pairs, whose brightness fluctuates in a pattern strikingly similar to an electrocardiogram. This automated approach marks a significant advance over the traditional method of manual inspection, which has been the primary way of finding these objects until now.
The new technique dramatically speeds up the search for these celestial oddities, which are characterized by two stars in a highly elliptical orbit. As the stars swing close to each other at their point of nearest approach, their mutual gravity causes them to stretch and deform, leading to a distinct, sharp surge in brightness. By training a neural network to recognize this unique light curve signature, researchers have created a powerful new tool for stellar classification that has already identified four previously unknown heartbeat stars from an existing catalog of eclipsing binaries.
Overcoming Manual Discovery
For decades, the discovery of heartbeat stars depended on the painstaking work of astronomers visually inspecting countless light curves from photometric surveys. A light curve is a graph that plots the brightness of a celestial object over time. The signature of a heartbeat star is a distinct brightening event that occurs periodically. However, the exact shape of this brightening can vary significantly, making it a challenging pattern to search for algorithmically.
This reliance on manual discovery was both time-consuming and inherently limited, meaning that many potential candidates in the vast datasets produced by modern telescopes were likely missed. The recent explosion in data from surveys like NASA’s Transiting Exoplanet Survey Satellite (TESS) and the earlier Kepler mission has made manual analysis increasingly impractical. Machine learning and AI offer a scalable solution, capable of analyzing millions of light curves far more efficiently than human researchers.
Training an AI for the Task
The research team developed a novel approach using a specific type of AI called a recurrent neural network (RNN), which is particularly effective at recognizing patterns in sequential data like time series. Instead of feeding the raw light curve data directly into the network, the scientists first devised a method to extract the most important features that characterize a heartbeat star’s signal.
Feature Extraction
The process begins by transforming the light curve data using a Fourier analysis, which breaks down the complex signal into a series of simpler, repeating frequencies. From this spectrum, the researchers calculated the star system’s orbital frequency and then extracted the amplitudes of the first 100 harmonics of that frequency. These 100 harmonic amplitudes were then normalized and used as a compact and powerful “feature vector” that describes the essential shape of the light curve. This method proved to be an effective way to represent the defining characteristics of a heartbeat star while significantly reducing the amount of data the neural network had to process.
Simulated Universe
To train the RNN, the team generated a large dataset of synthetic heartbeat star light curves using a software package called ELLC. This allowed them to create a wide variety of examples with known properties, such as orbital eccentricity. The network was then trained on these simulated feature vectors, learning to correlate specific harmonic patterns with the high eccentricities that define heartbeat stars.
Performance and Real-World Validation
The AI’s performance was first evaluated using a test set of synthetic light curves it had never seen before, where it achieved an impressive 95% accuracy in correctly identifying the systems. The true test, however, was applying the trained model to real-world observations. When tasked with analyzing the light curves of known heartbeat stars discovered by the OGLE, Kepler, and TESS surveys, the network maintained a high average accuracy of 86%.
Putting the system to practical use, the researchers applied their method to a catalog of eclipsing binary stars compiled by Kirk et al. From this dataset, the AI successfully identified four new heartbeat star systems that had not been previously classified as such. This result demonstrates the method’s effectiveness as a practical tool for new astronomical discoveries.
Why Heartbeat Stars Matter
Heartbeat stars are more than just a celestial curiosity; they are valuable natural laboratories for studying the effects of gravity and tides. The periodic close encounters between the two stars in a highly eccentric orbit cause immense tidal forces that are much stronger than those experienced by stars in nearly circular orbits. These forces stretch and distort the stars, an effect known as tidal distortion, which is the primary cause of the observed brightness variations.
By studying the precise shape of the light curve, astronomers can infer properties of the stars and their orbits, including the orbital period, eccentricity, and orientation in space. Furthermore, analyzing how the stellar material responds to these tidal forces provides crucial insights into the internal structure and physics of stars. In some systems, these powerful tidal interactions can even excite pulsations within the stars, offering another window into stellar astrophysics.
The Future of Automated Astronomy
The success of this recurrent neural network provides a new and efficient pathway for identifying heartbeat stars and other variable stellar objects. The methodology is not limited to this specific class of star; it can be adapted to recognize other types of periodic variable stars by training it on different sets of light curve features. This flexibility is crucial for maximizing the scientific return from the massive datasets generated by current and future sky surveys.
As astronomy enters an era of big data, with projects like the Vera C. Rubin Observatory poised to generate petabytes of information, AI-driven discovery will become indispensable. Automated systems like this one will allow researchers to quickly and accurately classify objects of interest, freeing up valuable human time for more detailed analysis and interpretation. This work represents a key step toward a future where AI and human astronomers work in tandem to explore the cosmos.