5:15 PM - 7:15 PM
[MGI30-P06] Machine Learning Model Predicting Planetary Formation in the Giant Impact Stage
Keywords:Planet formation, N-body simulation, Machine learning
There are two major problems in the current standard model of terrestrial planet formation. The first is the difficulty to explain the diversity in the distribution of observed exoplanets. The second is that most observed exoplanet systems are evolved, and only a few examples of systems in the formation process are available to validate the physical processes in planet formation models. To address these issues, planet population synthesis models have been developed (e.g., Ida & Lin 2004, Mordasini et al. 2009, Kimura & Ikoma 2020). These models accelerate the calculations of planet formation by adopting empirical formulas for each formation stage, and they can simulate the evolution of planetary systems from dust to the giant impact stage. The models enable statistical studies of planet formation from various initial conditions.
The giant impact stage in planet population synthesis models has been modeled using either empirical formulas or N-body simulations. However, using empirical formulas result in unrealistic results under some conditions (Kimura et al. in prep.), while N-body simulations provide higher accuracy but are expensive. Recently, there have been studies on replacing N-body simulations in the giant impact stage with machine learning models. Machine learning models can predict the final state of a planetary system with N-body simulation-like accuracy in four orders of magnitude less time (Lammers et al. 2024). However, the current accuracy of the machine learning model in predicting the pairs of colliding protoplanets is only about 60%, which is insufficient for studying the properties of planetary systems formed through the giant impact stage.
In this paper, we create neural networks to predict the pair of planets which will collide in the system. Our goal is to shorten computation time while maintaining the statistical accuracy of results by creating a machine learning model that predicts the orbital and accretionary evolution of planets based on the results of N-body simulations for the giant impact stage. This will enable the parameter surveys using the planet population synthesis model while maintaining the calculation accuracy of the giant impact stage, which allows the validation of physical processes in planet formation. Here we discuss in detail the creation process of the machine learning model and the prediction results.
The giant impact stage in planet population synthesis models has been modeled using either empirical formulas or N-body simulations. However, using empirical formulas result in unrealistic results under some conditions (Kimura et al. in prep.), while N-body simulations provide higher accuracy but are expensive. Recently, there have been studies on replacing N-body simulations in the giant impact stage with machine learning models. Machine learning models can predict the final state of a planetary system with N-body simulation-like accuracy in four orders of magnitude less time (Lammers et al. 2024). However, the current accuracy of the machine learning model in predicting the pairs of colliding protoplanets is only about 60%, which is insufficient for studying the properties of planetary systems formed through the giant impact stage.
In this paper, we create neural networks to predict the pair of planets which will collide in the system. Our goal is to shorten computation time while maintaining the statistical accuracy of results by creating a machine learning model that predicts the orbital and accretionary evolution of planets based on the results of N-body simulations for the giant impact stage. This will enable the parameter surveys using the planet population synthesis model while maintaining the calculation accuracy of the giant impact stage, which allows the validation of physical processes in planet formation. Here we discuss in detail the creation process of the machine learning model and the prediction results.