The development of reliable Digital Twins (DT) should include uncertainty quantification. However, DTs should also be efficient and run in real-time. Uncertainty propagation of a computational model is most commonly performed using a Monte Carlo (MC) simulations. To reduce the typically excessive computational cost associated with MC alternative methods requiring a smaller amount of samples can be used. The two most common ones are gaussian process regression (GPR) [1] and polynomial chaos expansions (PCE) [2]. This work suggests another alternative in the form of a neural network, in an effort to further reduce the computational cost. The proposed strategy will employ a neural network trained with data generated by Monte Carlo simulations of the computational model for different distributions of uncertain input parameters. The neural network input consists of distributions of the uncertain parameters and can be enhanced using extra samples of the quantity of interest from the computational model. Its output is an estimate of the distribution of the quantity of interest. Our study considers the turbulent flow around an airfoil, with the uncertainty residing in the Reynolds number. For the sake of computational efficiency, the computational model is replaced by its hybrid ROM counterpart. Comparison between a Monte Carlo simulation, PCE surrogate and the neural network is presented, while also discussing the error between the ROM and the underlying full order model.