Magnetic fields play a very important role in stellar evolution, as well as vary depending on the evolutionary stage. To understand how the stellar magnetic fields evolve is necessary to measure and map the magnetic fields over the stellar surface. It can be done through spectropolarimetric observations through the four Stokes parameters (I, Q, U, and V). In this work, we propose a deep-learning approach to estimate the Stokes parameters based on eight input parameters (dipolar moment strength, m; the magnetic dipole position inside the star, X-2, X-3; the rotation phase, p; the magnetic geometry of the dipolar configuration, a, ss, gamma; and the inclination angle of the stellar rotation axis with respect to the line of sight, i) and using a synthetic dataset generated by cossam. Different configurations of a neural network have been experimented with: the number of layers and neurons; the scaling of the input and output parameters; the size of training data; and estimating separately and jointly the output parameters. The best configuration of the neural network model scores a mean squared error of 1.4e-7, 2.4e-8, 1.5e-8, and 1.3e-7, for Stokes I, Q, U, and V, respectively. In summary, the model effectively estimated the Stokes I and V, which respectively correspond to the total intensity and circular polarization of the light emitted by magnetic stars; however, struggled with the Stokes Q and U, which represent linear polarization components generally for very small m. Overall, our work presents a promising avenue for advancing our understanding of stars that host a magnetic field.
