Can a Machine Learn the Outcome of Planetary Collisions?
Abstract
Planetary-scale collisions are common during the last stages of formation of solid planets, including the solar system terrestrial planets. The problem of growing planets has been divided into studying the gravitational interaction of embryos relevant on million year timescales and treated with N-body codes and the collision between objects with a timescale of hours to days and treated with smoothed-particle hydrodynamics. These are now being coupled with simple parameterized models. We set out to investigate if machine-learning techniques can offer a better solution by predicting the outcome of collisions that can then be used in N-body simulations. We considered three different supervised machine-learning approaches: gradient boosting regression trees, nested models, and Gaussian processes (GPs). We found that the former produced the best results, and that it was slightly surpassed by ensembling different algorithms. With GPs, we found the regions of parameter space that may yield the most information to machine-learning algorithms. Thus, we suggest new smoothed-particle hydrodynamics calculations to focus first on mass ratios ≳0.5.
- Publication:
-
The Astrophysical Journal
- Pub Date:
- September 2019
- DOI:
- 10.3847/1538-4357/ab2bfb
- arXiv:
- arXiv:1902.04052
- Bibcode:
- 2019ApJ...882...35V
- Keywords:
-
- planets and satellites: formation;
- Astrophysics - Earth and Planetary Astrophysics
- E-Print:
- 13 pages, 11 figure, 2 tables. Submitted to ApJ