Paper detail

Realistic On-the-fly Outcomes of Planetary Collisions II: Bringing Machine Learning to N-body Simulations

Terrestrial planet formation theory is at a bottleneck, with the growing realization that pairwise collisions are treated far too simply. Here, and in our companion paper (Cambioni et al. 2019) that introduces the training methodology, we demonstrate the first application of machine learning to more realistically model the late stage of planet formation by giant impacts. We present surrogate models that give fast, reliable answers for the masses and velocities of the two largest remnants of a giant impact, as a function of the colliding masses and their impact velocity and angle, with the caveat that our training data do not yet include pre-impact rotation or variable thermal conditions. We compare canonical N-body scenarios of terrestrial planet formation assuming perfect merger (Chambers 2001) with our more realistic treatment that includes inefficient accretions and hit-and-run collisions. The result is a protracted tail of final events lasting ~200 Myr, and the conversion of about half the mass of the initial population to debris. We obtain profoundly different solar system architectures, featuring a much wider range of terrestrial planet masses and enhanced compositional diversity.