SpecFANN: Spectral Fitting via Artificial Neural Networks. I. A deep-learning-based emulator and fitting suite

The importance of massive stars cannot be overstated: they are powerful probes of the early universe, they play a vital role in the chemical and mechanical evolution of their host environments, and their end products allow us to study the most extreme physics in the universe. Obtaining accurate stellar and surface parameters for large samples of massive stars is vital to our understanding of how they evolve and how their births, lives, and deaths affect their surroundings. With the large volume of data expected from the next generation of spectroscopic surveys, it is likely that the computational cost of our current analysis methods (especially the calculation of model stellar atmospheres and synthetic spectra) will prove to be the most important bottleneck impeding our progress. In order to overcome these limitations and dramatically decrease computing times, we aim to develop a robust emulator for the radiative transfer and spectral synthesis code. Additionally, we aim to explore alternative fitting methods that have not been feasible up to this point due to computational costs. We calculated a large set (∼ 50,000 models in total) of synthetic spectra of OB-type stars, varying temperature, surface gravity, helium mass fractions, and CNOSi abundances, and we trained a collection of neural networks to emulate these models with separate networks for each line. We also developed the open-source python package (Spectral Fitting via Artificial Neural Networks), which provides users with a suite of fitting methods that can be used with these or other user-generated neural networks. The majority of the trained neural networks reach average accuracies of better than ∼0.01-0.1% for photospheric lines and better than ∼0.1-1% for wind lines. We find that is able to obtain robust and accurate stellar parameters that are consistent with the literature for a sample of 52 early-type stars. Using we find that we can achieve the same fit in sim1/360,000 of the time when compared to alternative techniques that rely on "on-the-fly" computations. We have demonstrated that neural networks offer a viable path forward to address the computational limitations of our current atmosphere analysis and stellar parameter determination methods for hot stars.