Accurate Teff and [M/H] determinations for the CARMENES M dwarfs from deep transfer learning

Accurate Teff and [M/H] determinations for the CARMENES M dwarfs from deep transfer learning We present a new technique for deriving photospheric stellar parameters of M dwarfs with high accuracy, the so-called transfer learning. There, information from one domain is being transferred to improve the accuracy of a neural network model in another domain. We demonstrate the feasibility of the deep transfer learning approach applied to M-dwarf high-resolution spectra from the CARMENES survey. In order to do so, we use 14 stars of the CARMENES sample with interferometric angular diameters to calculate the effective temperature, as well as six M dwarfs being common proper-motion companions to FGK-type primaries with known metallicity. After we train a deep learning neural network model on synthetic PHOENIX-ACES spectra, we use the internal feature representations together with those 14+6 independently labeled stars as a new input for the transfer process. We compare the derived stellar parameters of a small sample of M dwarfs kept out of the training phase with results from other methods. We demonstrate that deep transfer learning provides a higher accuracy than our previous deep learning method, indicating that it is a robust tool for obtaining M-dwarf stellar parameters when measured against samples from independent estimations for well known stars