An Interpretable Machine-learning Framework for Modeling High-resolution Spectroscopic Data*
چکیده
Comparison of echelle spectra to synthetic models has become a computational statistics challenge, with over ten thousand individual spectral lines affecting typical cool star spectrum. Telluric artifacts, imperfect line lists, inexact continuum placement, and inflexible frustrate the scientific promise these information-rich datasets. Here we debut an interpretable machine-learning framework "blas\'e" that addresses other challenges. The semi-empirical approach can be viewed as "transfer learning" -- first pre-training on noise-free precomputed models, then learning corrections depths widths from whole-spectrum fitting observed auto-differentiable model employs back-propagation, fundamental algorithm empowering modern Deep Learning Neural Networks. Here, however, 40,000+ parameters symbolize physically profile properties such amplitude, width, location, shape, plus radial velocity rotational broadening. This hybrid data-/model- driven allows joint modeling stellar telluric simultaneously, potentially transformative step forwards for mitigating deleterious contamination in near-infrared. blas\'e acts both deconvolution tool model. general purpose scaffolding may extensible many applications, including precision velocities, Doppler imaging, chemical abundances, remote sensing. Its sparse-matrix architecture GPU-acceleration make fast. open-source PyTorch-based code includes tutorials, Application Programming Interface (API) documentation, more. We show how fits into existing Python spectroscopy ecosystem, demonstrate range astrophysical discuss limitations future extensions.