M DWARF STELLAR PARAMETER DETERMINATION WITH AUTOENCODERS AND DEEP TRANSFER LEARNING - PEDRO MAS BUITRAGO

The estimation of stellar parameters for M dwarfs often involves the comparison of observed spectra with different synthetic collections. In this process, a major source of uncertainty is the “synthetic gap” (difference between theoretical and observed spectra), which must be addressed beforehand to know the reliability of the parameter estimation.
In this work, we propose a deep learning (DL) methodology to bridge the synthetic gap in stellar parameter estimation, using a sample of high S/N, high resolution, spectra from 286 CARMENES survey M dwarfs. For this purpose, we built a two-step process that involves different deep learning approaches. In particular, we used deep transfer learning (DTL), which focuses on transferring knowledge from one model to another.
First, we trained a sparse autoencoder to effectively compress synthetic spectra from the PHOENIX-ACES models into a low-dimensional latent space. Using this trained autoencoder as a base model, we adopted a DTL approach to adapt it to the target domain (CARMENES spectra) while keeping the features learned in the data-rich source domain (PHOENIX-ACES spectra) frozen.
Using the low-dimensional encoded latent space from the PHOENIX-ACES spectra as input features, we trained a convolutional neural network (CNN) to build a regression model and estimate the stellar parameters of the 286 M Dwarfs.