Year
2023
Location
UCSD Astrophysics
Category
Research
Duration
1 Year
Ultracool dwarfs (UCDs), including spectral types late-M through Y, are very low-mass stars and brown dwarfs (M ≤ 0.1 M⊙). Some exist in binary systems, which tend to have very small separations (~4–7 AU), making them difficult to resolve visually.
Spectral binaries, unresolved systems detected via peculiarities in combined-light spectra, have traditionally been identified using spectral indices (Burgasser et al. 2010; Bardalez Gagliuffi et al. 2014), though these indices also pick up non-binaries and have unclear biases.
This study explores the use of random forest (RF) models as a supervised machine learning method for classifying UCD spectral binaries.
Data: 414 spectra from the SpeX Prism Library with S/N > 5.
Preparation: Spectra reclassified, interpolated, flux-calibrated, and normalized at 1.2 μm.
Binary templates were created by pairing M7–L7 primaries with T1–T8 secondaries. Flux uncertainties were used to simulate measurement noise.
Features: Normalized flux values across 0.9–2.4 μm.
Class imbalance: Handled by oversampling using Monte Carlo simulations to generate ~100,000 single and binary templates.
Four RF models (50 decision trees each) were trained across S/N bins (0–50 to 150–200), using a 75/25 train-validation split.
Precision: Ranged from 95%–100%, dropping to 87% only in the lowest S/N bin (<10).
Comparison to prior indices: RF models outperformed index-based methods (which had 72% precision for singles and 89% for binaries).
Feature importance: Key discriminatory regions largely overlapped with those in prior work, but RF discovered new features (notably at 1.18 μm) not previously used.
The pilot study shows that random forest models significantly improve binary classification accuracy over traditional spectral index approaches, even identifying previously unknown discriminative spectral features.
Future work will expand the method to a broader set of UCD combinations, extract component types, and apply the trained model to large-scale surveys to discover new UCD binaries.
