Digital twins, the virtual representation of physical systems, are experiencing increasing popularity due to their transformative role in design, optimization and decision making. The two main components for digital twins are data and simulators [1]. While reduced order models (ROMs) have been shown to offer excellent speed-up for the approximation of complex physical problems, constructing ROMs for large problems is often infeasible. Moreover, the accuracy of the reduced model is measured, often inadequately, in terms of how close it can get to the underlying high-fidelity model, instead of real-life data. This work will use the hybrid analysis and modeling [2] (HAM) strategy combining multiple reduced order modeling (ROM) techniques and data-driven neural network-based corrections. As mentioned, this work will initially consider the construction of a hybrid ROM scheme [3], combining the proper orthogonal decomposition (POD) with Galerkin projection and interpolation schemes. Aiming towards efficiency and coarse reduced bases, we use HAM, and more specifically the corrective source term approach (CoSTA). We introduce an extra correction step in the typical ROM online problem that uses a meta-model to evaluate the residual error of the ROM approximation, modify the algebraic problem and improve the final solution. Finally, the ability of the proposed strategy is tested in terms of computational acceleration and increase in accuracy when approximating turbulent flow around known wing geometries.