On October 14, in a significant breakthrough for astronomy, researchers from the Chinese Academy of Sciences announced the discovery of five new exoplanets that are smaller than Earth and have orbital periods of less than one day. This exciting development comes in the wake of the identification of ultrashort-period exoplanets, which has posed unique challenges and opportunities for existing theories of planetary formation since their initial discovery in 2011 through data from the Kepler Space Telescope.
Leading this international team, Professor Ge Jian from the Shanghai Astronomical Observatory utilized an innovative deep learning algorithm to analyze stellar data released by Kepler in 2017. Among this research, four of the newly identified planets are the smallest known exoplanets in close proximity to their solar-type stars, with sizes comparable to Mars.
“This landmark achievement marks the first instance where astronomers have harnessed artificial intelligence to conduct a comprehensive search and identification of potential signals from exoplanets in a single stroke,” Ge explained. The results of their study were recently published in the Monthly Notices of the Royal Astronomical Society.
When asked about the advantages of the new algorithm, Ge elaborated, “After continuous innovation, our team successfully developed a new algorithm that integrates phase folding with deep learning via convolutional neural networks. This approach improves search speed by approximately 15 times compared to commonly used algorithms and enhances both detection accuracy and completeness by about 7%. The new method demonstrates significant potential in identifying faint transit signals.”
The algorithm was effectively applied to the Kepler dataset, successfully identifying five ultrashort-period planets designated as Kepler—158d, Kepler—963c, Kepler—879c, Kepler—1489c, and Kepler—2003b. Notably, Kepler—879c and others rank among the smallest found in the category of ultrashort-period planets, all within five stellar radii of their host stars.
Discussing the implications of these findings, the research team noted that the Mars-sized planets add diversity to the sample of known exoplanets and provide fresh insights into the mechanisms of forming ultrashort-period planets. Their existence lays an important foundation for understanding early planetary evolution, as well as dynamics concerning planetary interactions and stellar influences.
“This work represents a milestone in the application of AI in astronomical big data,” Ge stated. To achieve such discoveries, he emphasized the need for innovative AI algorithms and vast synthetic datasets to train them effectively, aimed at swiftly and accurately detecting faint signals that traditional methods might miss.
Professor Josh Winn, an astrophysicist from Princeton University, shared his enthusiasm: “Ultrashort-period planets, often termed ‘lava worlds,’ exhibit extreme and unexpected characteristics that offer clues about the orbital changes of planets over time. I initially thought we had exhausted the Kepler data for transit signals, so hearing about these new potential discoveries is thrilling, and I’m impressed by this technological accomplishment in discovering new planets.”
The research team pointed out that ultrashort-period planets are quite rare, occurring with a frequency of about 0.5% around solar-type stars. To date, astronomers have confirmed only 145 ultrashort-period planets, with merely 30 smaller than Earth. Understanding their relative abundance and characteristics is critical for validating theoretical models, yet the limited sample size makes it challenging to glean accurate statistical insights.
Reflecting on the research’s origin, Ge mentioned, “The journey truly began in 2015, inspired by the breakthroughs of AlphaGo in defeating professional players at Go. Motivated by my colleague Professor Li Xiaolin from the University of Florida, I sought to apply deep learning to Kepler’s light curve data to find transit signals that traditional methods had overlooked. Fortunately, after nearly a decade of effort, we are now celebrating our first significant results.”
The research implemented a rapid folding algorithm optimized for GPUs to enhance low signal-to-noise ratio transit signals, enabling high-precision searches. However, the scarcity of known true transit signals posed a challenge for training the neural network effectively. To overcome this, the team innovated by designing various potential transit signal templates based on physical characteristics and trained the network using a large dataset composed of over 2 million synthesized light curves generated from Kepler’s real data. The trained neural network was then applied to the Kepler dataset alongside the fast folding algorithm, successfully leading to the discovery of five small-radius ultrashort-period exoplanets.
