Transfer learning technique applied to polarized spectral lines

We present a new approach to produce polarized spectra in the four Stokes profiles using a deep artificial neural network (ANN) model. The novelty of our approach lies in the use of the so-called transfer learning technique, where a pre-trained ANN is used to train other ANNs to perform similar tasks – such as the synthesis of Stokes profiles at different wavelengths and/or with a different atmospheric model. The training of the original ANN required millions of instances, each corresponding to different magnetic configurations over the stellar surface. This process can be very time-consuming, particularly when analysing several spectral lines. However, once the original ANN is trained, the transfer learning technique allows subsequent ANNs to achieve comparable performance with only a few thousand instances. This approach has the significant advantage that many ANNs can be trained with small data bases, which is a key aspect in machine learning. We demonstrate the performance of our approach through a test study involving the inversion of multiline profiles, finding excellent results in determining the stellar surface magnetic configuration.