Autoencoders (AEs) are popular in Fluid Mechanics applications for the construction of non-linear reduced order models of complex physical phenomena. Collecting high-fidelity data for training AEs (and, in general, deep learning algorithms) is computationally expensive, and there is a growing interest in multi-fidelity approaches for AEs [1]. We propose a methodology for the construction of autoencoders able to provide high-fidelity predictions with quantified uncertainties with low computational efforts (taking advantage from multi-fidelity methods) on a problem of practical engineering interest. We introduce a Multi-Branch Autoencoder (MBA) architecture that can manage dataset composed of two-dimensional data with channels not spatially related. The model is trained using a multi-fidelity algorithm recently developed in [2], which uses a database composed by many low- and few high-fidelity simulations, substantially reducing the computational cost of database construction. An ensemble of trained models is built to provide predictions with quantified uncertainties. We applied the present method on a database composed of RANS-LES nozzle jets with different geometries, varying the shape of the outlet section and the distribution of the cross-section area. Figure 1a shows a typical prediction (with standard deviation) of the model and 1b a comparison between the autoencoder predictions and the reference LES results in terms of streamwise velocity, highlighting the good accuracy of the autoencoder. A significant advantage of this method is the ability to capture the statistical flowfield with a low number of time steps for the LES time-averages, as will be shown.