Machine learning applied to simulations of collisions between rotating, differentiated planets

Area of Science:

• Planetary Science
• Astrophysics
• Computational Science

Background:

• Planetary collisions are crucial for terrestrial planet formation.
• Current methods for simulating collisions lack accuracy and scope.
• Advancements in machine learning offer new approaches.

Purpose of the Study:

• To develop and evaluate data-driven techniques for predicting planetary collision outcomes.
• To compare machine learning methods with existing analytical and semi-analytical approaches.
• To demonstrate the generalizability of these techniques for various post-impact quantities.

Main Methods:

• Utilized 14,856 SPH simulations of pairwise collisions between differentiated bodies.
• Trained and evaluated four data-driven techniques: ensemble methods, neural networks, Gaussian processes, and polynomial chaos expansion.
• Compared performance against established analytical and semi-analytical models.

Main Results:

• Data-driven emulators achieved high accuracy in classifying and predicting collision outcomes.
• Machine learning methods demonstrated superior performance compared to existing techniques.
• The developed emulators are generalizable to diverse post-impact properties.

Conclusions:

• Data-driven emulators can accurately predict planetary collision outcomes.
• These emulators are poised to enhance N-body simulations for planet formation studies.
• This approach offers a more efficient and accurate alternative to traditional simulation methods.