Multi-Output Random Forest Regression to Emulate the Earliest Stages of Planet Formation

TL;DR

This paper introduces a multi-output random forest regression model that accurately and efficiently approximates dust particle size distributions in early planet formation, significantly reducing computational costs compared to traditional brute-force simulations.

Contribution

The paper presents a novel machine learning approach using multi-output random forests to emulate complex coagulation equations in planet formation, enabling faster simulations.

Findings

Achieved an R² of 0.97 in predictions.

Generated results significantly faster than brute-force methods.

Provided a scalable approach for simulating early planet formation processes.

Abstract

In the current paradigm of planet formation research, it is believed that the first step to forming massive bodies (such as asteroids and planets) requires that small interstellar dust grains floating through space collide with each other and grow to larger sizes. The initial formation of these pebbles is governed by an integro-differential equation known as the Smoluchowski coagulation equation, to which analytical solutions are intractable for all but the simplest possible scenarios. While brute-force methods of approximation have been developed, they are computationally costly, currently making it infeasible to simulate this process including other physical processes relevant to planet formation, and across the very large range of scales on which it occurs. In this paper, we take a machine learning approach to designing a system for a much faster approximation. We develop a multi-output random forest regression model trained on brute-force simulation data to approximate distributions of dust particle sizes in protoplanetary disks at different points in time. The performance of our random forest model is measured against the existing brute-force models, which are the standard for realistic simulations. Results indicate that the random forest model can generate highly accurate predictions relative to the brute-force simulation results, with an $R^{2}$ of 0.97, and do so significantly faster than brute-force methods.