Speaker
Description
The outflows of massive stars significantly affect their stellar evolution and surroundings. The mass-loss rates of these stars is thus essential to constrain from their stellar spectra. However, this requires the detailed spectroscopic analysis of large samples of stars. Precisely modelling the wind and atmospheres of massive stars is computationally very expensive, which severely limits the sample of stars that can be analysed. By replacing a 1D spectral atmosphere code with a neural network as a surrogate model, generating synthetic spectra becomes several thousand times faster, allowing the use of previously computationally unfeasible but statistically more robust sampling algorithms such as using Markov Chain Monte Carlo or simulation-based inference. In this talk I'll discuss our recent development of a neural network emulator for the radiative transfer code FASTWIND, and discuss how well neural networks perform when predicting stellar spectra of LMC O-stars. I will give insights into how their performance compares to using stellar atmosphere codes when inferring the parameters of observed stellar spectra. I will also compare the already established optimization methods to previously inaccessible sampling methods.
| Talk category | Splinter 6: Data science |
|---|---|
| Second preference | NOVA Network 2 |
| PhD relevance | 1st |
