Â鶹ÒùÔº

Over the course of more than two decades, researchers at the University of Bern have developed the so-called "Bern model," a suite of computer programs that can numerically simulate the formation of planetary systems, thus shedding light on system architecture. These models are, however, very complex: each simulation from the Bern model can take a few days to a few weeks to be computed using modern supercomputers.

Using modern AI techniques trained on the Bern model data, Prof. Yann Alibert and Sara Marques from the NCCR PlanetS and the Center for Space and Habitability of the University of Bern, and Dr. Jeanne Davoult, former Ph.D. student of the University of Bern and now researcher at the DLR in Berlin, have developed an AI model capable of computing the formation of planetary systems in seconds, a million times faster than traditional computations.

The study has just been in the journal Astronomy and Astrophysics and was presented last week at the "Fast Machine Learning for Science" conference in Zurich (where it won the prize of the best poster) and this week at the Joint Meeting of the Europlanet Science Congress and the Division for Planetary Sciences () 2025 in Helsinki.

Knowing where to observe

Present day and near future observational facilities will soon be able to observe and characterize extrasolar planets similar to Earth, while they so far have been limited to planets closer to their host stars. "Earth-like planet detection requires large amounts of observing time. In this context, knowing where to observe is very important to save very costly observation time," explains Yann Alibert, first author of the study.

In order to prioritize between different possible targets, one can use the observations of easier-to-observe other planets in the same systems. This, however, requires a profound understanding of the so-called architecture of a system: how the properties (orbital position, mass, etc.) of one planet in a system relate to the properties of other planets in the same system.

Inspired by large language models

The team trained its AI model on tens of thousands of numerical simulations of planetary system formation also developed at the University of Bern. "The new AI model can be used to predict the presence and properties of yet-to-be-discovered additional planets in already known extrasolar planetary systems," as Sara Marques, Ph.D. student at the University of Bern, points out.

In an experiment presented in the current study, the authors showed that in a real three-planet system, the properties of the second and third planet can be inferred from the properties of the innermost planet of the system.

Alibert explains, "This approach can be used to generate new planetary systems: Knowing a single planet in a system, we can predict the rest of the planets for systems of three planets with our model.

"The key in our study was to realize that planetary systems can be seen as sequences of planets, exactly as sentences are sequences of words. This triggered the idea of using the AI methods from large language models, used for instance by chatbots such as ChatGPT, to build our AI model."

The authors used the so-called "Transformer architecture" introduced in the field in 2017 to create a generative model that can produce sequences of planets orbiting the same stars. "The large language models predict the rest of a sentence based on the sequence created by the first few words. In our case, we predict the sequence of outer planets in a system, based on the first inner ones," further explains Marques.

"This new study builds upon a previous AI model encouraging results," points out Dr. Davoult, former student at the NCCR PlanetS, now working at the DLR Berlin.

"In the last model, based on the inner planet of a system, we were predicting the probability of an Earth-like planet to be in the system. Keeping the analogy with language models, it was like predicting the presence of a specific word in a sentence, based on its beginning. In this new study, we predict all the rest of the sentence and not only the probability of a single word."

"The results of the generative AI model were so accurate that we were very skeptical at first," remembers Marques. A large range of tests were conducted by the researchers, in which they used machine learning classifiers, and they submitted their results to other scientists. "In the end, they all concluded the same: generated planetary systems are virtually indistinguishable from numerical simulations."

Preparing for the PLATO mission and others

Scheduled to be launched in 2026, the ESA PLATO mission will discover thousands of planetary systems, with the planet closest to the star being, in general, the first to be observed. Some of these systems could harbor planets like Earth, yet these will likely be discovered by ground-based telescope using other observations later.

"Our new AI model could be used to prioritize the observations of these systems by telescope, enhancing the probability to find Earth twins," says Davoult. In the coming years, the models will be extended to predict more properties of planets, such as their composition or habitability.

"When I was hired as a postdoc in 2001, I initiated numerical simulations of planetary systems at the University of Bern. This new AI model is the natural continuation of this Bernese expertise," says Alibert. "AI is now present in everyone's life. I am convinced it will be more key to scientific discoveries, in planetary sciences and elsewhere."