TelescopeML. II. Convolutional Neural Networks for Predicting Brown Dwarf Atmospheric Parameters

Abstract

Accurately and swiftly predicting the parameters of brown dwarf atmospheres from observational spectra is crucial for understanding their atmospheric composition and guiding future follow-up observations. Here, we utilized convolutional neural networks (CNNs) as a high-performance deep learning algorithm, training them with synthetic spectra generated from Sonora–Bobcat cloudless and equilibrium model grids to infer effective temperatures ( T _eff ), surface gravities ( $\mathrm{log}\,g$ ), carbon-to-oxygen ratios (C/O), and metallicities ([M/H]). The performance metric, or R ^2 value (the coefficient of determination), of the trained CNNs in predicting T _eff , $\mathrm{log}\,g$ , C/O, and [M/H] is 1.0, 1.0, 0.99, and 0.93 on both the training and test sets, respectively. We apply our trained CNN model on the three well-known T dwarf benchmarks, HD 3651B, GJ 570D, and Ross 458C. For HD 3651B and GJ 570D, we find that our predicted T _eff and $\mathrm{log}\,g$ are in good agreement with some of the previous modeling studies from forward and retrieval approaches, but the [M/H] is underestimated up to ≈0.3 dex. For Ross 458C, our CNN model predicts the $\mathrm{log}\,g$ accurately but tends to overestimate T _eff by ≈150 K and [M/H] by ≈1.0 dex compared to the previous modelings, which likely happens because our model grids lack disequilibrium chemistry and clouds. We also find that the C/O ratio tends to be higher than solar if not kept fixed. Concurrently, all modules and functions used to train, develop, and deploy the CNNs and process the spectra are consolidated into a Python PyPI package named TelescopeML , which is publicly available along with a thoroughly documented tutorial covering installation, deployment, and core concepts.

Keywords

• Brown dwarfs
• Convolutional neural networks
• T dwarfs
• Atmospheric composition