|
Stellar spectra emulators often reach a limit in accuracy, leading to systematic errors in the estimation of stellar atmosphere parameters. I would like to introduce the TransformerPayne emulator and compare its performance with The Payne emulator (a fully connected neural network), its large version, and a convolutional-based emulator. When tested on synthetic spectra grids, TransformerPayne demonstrated superior performance, achieving a mean absolute error of approximately 0.15%, which is several times better than the other methods considered. Additionally, TransformerPayne excels when using a fine-tuning approach, allowing pretraining on grids of stellar spectra with simplified physics and transferring this knowledge to more complex grids, which reduces the required training grid size by up to tenfold. Furthermore, we show that an analysis of the emulator’s internal workings reveals interpretable features shared across spectral lines, offering a partial explanation for its enhanced accuracy and data efficiency. |
|