Supervised Machine Learning for Intercomparison of Model Grids of Brown Dwarfs: Application to GJ 570D and the Epsilon Indi B Binary System

Abstract

Self-consistent model grids of brown dwarfs involve complex physics and chemistry, and are often computed using proprietary computer codes, making it challenging to identify the reasons for discrepancies between model and data as well as between the models produced by different research groups. In the current study, we demonstrate a novel method for analyzing brown dwarf spectra, which combines the use of the Sonora, AMES-Cond and HELIOS model grids with the supervised machine learning method of the random forest. Besides performing atmospheric retrieval, the random forest enables information content analysis of the three model grids as a natural outcome of the method, both individually on each grid and by comparing the grids against one another, via computing large suites of mock retrievals. Our analysis reveals that the different choices made in modelling the alkali line shapes hinder the use of the alkali lines as gravity indicators. Nevertheless, the spectrum longward of 1.2 micron encodes enough information on the surface gravity to allow its inference from retrieval. Temperature may be accurately and precisely inferred independent of the choice of model grid, but not the surface gravity. We apply random forest retrieval to three objects: the benchmark T7.5 brown dwarf GJ 570D; and Epsilon Indi Ba (T1.5 brown dwarf) and Bb (T6 brown dwarf), which are part of a binary system and have measured dynamical masses. For GJ 570D, the inferred effective temperature and surface gravity are consistent with previous studies. For Epsilon Indi Ba and Bb, the inferred surface gravities are broadly consistent with the values informed by the dynamical masses.