Astronomers are focused on discovering planets that closely resemble Earth in size, composition, and temperature. Earth-like planets face numerous challenges in this quest. These planets are small and rocky, making them hard to detect. The current methods of planet hunting tend to favor gas giants, complicating matters. For a planet to have temperatures similar to Earth, it must orbit its host star at a similar distance, similar to Earth’s orbit around the Sun. This means it takes about a year to complete its orbit around the star. This raises an additional challenge for astronomers: locating Earth-like planets around a star requires telescopes to be dedicated to monitoring them for more than a year.
To maximize efficiency and reduce time spent on monitoring, scientists are seeking alternative methods to identify promising stars for in-depth searches before committing resources. A team of astronomers explored whether observable characteristics of planetary systems could indicate the presence of Earth-like planets. They found that the arrangement of known planets, along with their mass, radius, and proximity to their nearest star, could help predict the likelihood of Earth-like planets existing in those systems.
How effectively did the team test their approach using Machine Learning? They initiated their study by compiling a sample of planetary systems, some with Earth-like planets and some without. Since astronomers have only discovered about 5,000 stars that host orbiting planets, this sample size was too small for training machine learning models effectively. Consequently, the team generated three sets of planetary systems using a computational framework that simulates how planets form, based on the Bern model.
The Bern model initiates with 20 dust clumps, measuring around 600 meters, which is approximately 2,000 feet. These clumps help kickstart the accumulation of gas and dust into full-sized planets over a timespan of 20 million years. The planetary system evolves to a stable state over more than 10 billion years, leading to a Synthetic Planetary System that astronomers can utilize in their datasets. Using this model, they created 24,365 systems with sun-sized stars, 14,559 systems with similar stars, and 14,958 systems with different types of stars. Each group was further subdivided into those containing Earth-like planets and those without.
With these larger datasets in hand, the team utilized machine learning techniques known as Random Forest Models to categorize planetary systems based on their potential to host Earth-like planets. In a random forest setup, outputs are determined as either true or false through various components called trees that outline subsections of the entire training dataset. The team concluded that if a planetary system could host one or more Earth-like planets, the Random Forest algorithm should categorize it as “true.” They evaluated the algorithm’s accuracy using a metric known as the Precision Score.
The random forests made decisions based on specific characteristics within each synthetic planetary system. These factors included the number of planets, the presence of similar systems observed by astronomers, the system’s total planet count, and the mass and distance of planets over 100 times that of Earth, as well as the characteristics of the stars involved. The team allocated 80% of the synthetic planetary systems for training data, reserving the remaining 20% for initial testing of the completed algorithm.
The findings revealed that the random forest models accurately predicted where Earth-like planets are likely to exist with an impressive precision score of 0.99. Building on this success, they tested the model against data from 1,567 stars of similar sizes, each with at least one known orbiting planet. Out of these, 44 met the algorithm’s threshold for having Earth-like planets, suggesting that the majority of systems in this subset are stable enough to host such planets.
The team concluded that their models can effectively identify candidate stars for hosting Earth-like planets; however, they issued a caution. One concern is that the synthesis of planetary systems is time-consuming and resource-intensive, limiting the availability of training data. A more significant caution is rooted in the assumption that the Bern model accurately simulates the layered structure of planets. They urged researchers to rigorously validate their models for future theoretical work.
Post view: 230
Source: sciworthy.com
