AI Identifies 44 Previously Hidden Earth-like Planet Candidates
Jakarta, CNBC Indonesia — Artificial intelligence (AI) is beginning to help astronomers search for hidden Earth-like planets.
A machine-learning model developed by researchers at the University of Bern and Switzerland’s National Centre of Competence in Research (NCCR) Planet identified at least 44 planetary systems that could contain undiscovered Earth-like planets.
The study, published in the journal Astronomy & Astrophysics, found that the model achieved precision of up to 99% when tested on computer-simulated planetary systems.
However, the predicted planets around actual stars have not yet been confirmed by telescope observations. The method’s success still depends on further observations.
The AI model uses the arrangements of known planets to look for clues that other, unseen planets may exist. A planet’s mass and orbit can preserve traces of how an entire planetary system formed and evolved, including evidence of planets too small or faint for telescopes to detect.
Planets form together in discs of gas and solid material surrounding young stars. As they grow, migrate and interact with one another, these processes leave patterns in their masses and orbital distances.
The researchers tested whether those patterns could help distinguish systems containing small, temperate planets from those without them. Some systems have planets of similar sizes and at similar distances, a pattern often described as “peas in a pod”, while others have more varied arrangements.
Jeanne Davoult developed the machine-learning algorithm during her doctoral research at the University of Bern, working with Romain Eltschinger and Yann Alibert.
The researchers found that the strongest clues came from the observable architecture of a system, as well as the mass and orbital period of its innermost detected planet.
The team faced a challenge in training the model on observed planetary systems: astronomers rarely know all the planets orbiting a star. Smaller planets with longer orbits are also harder to find, leaving the picture obtained from observations incomplete.
The team therefore used the Bern Model of Planet Formation and Evolution to generate tens of thousands of synthetic planetary systems. The simulations covered stars with masses equivalent to that of the Sun, half the Sun’s mass and one-fifth of the Sun’s mass.
In these simulations, every planet was known, including those that observers might not be able to detect. The calculations began with planetary embryos inside discs of gas and small solid particles.
The researchers then simulated growth, migration and gravitational interactions over 20 million years. Subsequent calculations tracked cooling, contraction, atmospheric loss and tidal migration for up to 10 billion years.
Before training the model, the team concealed planets whose gravitational effect on their star fell below a set detection threshold. This was intended to reproduce the limitations of the radial-velocity method, which measures a star’s movement caused by an orbiting planet.
The model used a random-forest method comprising 500 decision trees. Each tree cast a vote on whether a system contained an Earth-like planet.
The researchers used 80% of the synthetic data to train the model and the remaining 20% to test it.
They prioritised precision: the proportion of systems that actually contain an Earth-like planet among all those the model classified as having one.
When the voting threshold was set above 90%, precision reached 99% for models trained on the two populations of higher-mass stars. The model for the lowest-mass stars achieved precision of 94%.