A Large SPH Collision Database and Machine Learning Framework for Predicting Post-Impact Remnants: Implications for Super-Mercury Formation

Abstract We present a large database of ~20, 000 smoothed particle hydrodynamics simulations of giant impacts aimed at understanding the formation of super-Mercuries. The dataset spans a broad parameter space that includes target masses from 0.01 to 12 M⊕, mass ratios 0.025 ≤ γ ≤ 1.0, impact parameters b ≤ 0.9, impact velocities from 1 to 25 Vesc, and varying iron core fractions and rotation rates. Using this data, we develop machine-learning models to predict the mass and iron mass fraction of post-collision remnants. Our analysis reveals distinct “sweet-spot” regions in the collision parameter space where remnants attain both significant mass and enhanced iron content. These regions are confined to collisions with impact velocities exceeding twice the mutual escape velocity and typically involve bodies of comparable mass. Applying a machine learning framework, we performed Monte Carlo simulations of both single collisions and collision chains to quantify the probability of super-Mercury formation across a range of impact conditions. We demonstrate that super-Mercury candidates with masses below and above approximately 5 M⊕ likely experienced distinct collision histories. In particular, for super-Mercuries with masses exceeding 5 M⊕ and iron mass fractions above 0.5, our analysis suggests that a series of collisions or a head-on impact is a more likely formation mechanism than a single oblique giant impact. Our analysis further reveals that the probability of super-Mercury formation varies significantly with the number of collisions in a chain, demonstrating that collision multiplicity—not only impact velocity—is a critical factor in modelling dense planet formation through sequential impacts.