Recovering the polarized cosmic microwave background (CMB) is essential for shedding light on the exponential expansion of the very early Universe, known as cosmic inflation. Achieving this goal requires not only improved instrumental sensitivity but also the development of robust and diverse data analysis techniques. In this work, we explore a novel component separation approach based on neural networks, previously validated using realistic Planck temperature simulations, to reconstruct the Stokes Q and U polarization maps. To validate the method, we first test the network on realistic Planck sky simulations of regions deliberately excluded from the training set. We compare the input and output EE and BB power spectra, finding a mean absolute error of 0.1 ± 0.3 μK 2 for the E-mode and -0.1 ± 0.3 μK 2 for the B-mode. These results demonstrate a partial recovery of the E-mode and a limited recovery of the B-mode, the latter remaining dominated by residual Planck noise. We then apply the trained network to public Planck observations, recovering CMB polarization maps broadly consistent with those obtained using the Commander method. The recovered EE spectra differ by less than 5% from the reference at intermediate and small angular scales, although significant discrepancies remain at large scales, which may impact cosmological interpretations. These results, while encouraging, clearly reflect the limitations of the current setup and motivate further improvements in training data and methodology. Based on these findings, we conclude that neural network-based methods show potential as component separation techniques in polarization CMB experiments. However, substantial improvements and more comprehensive analyses are necessary before these methods can provide reliable high-precision cosmological estimates.