Numerical simulation in fluid mechanics is extensively employed in industrial and academia for the prediction and resolution of complex fluid flow phenomena. However, in the case of high-fidelity numerical simulations, which are performed for studying complex phenomena, may require longer simulation times with high associated computational costs. Thus, there is a need to develop new robust techniques to reduce simulation time, while preserving the accuracy of results. Machine learning models have already been applied to fluid mechanics problems in order to predict the flow dynamics in a lower computational time, e.g., by performing a forecast about the future evolution of the dynamics [1] or by solving the differential equations [2]. In any case, ML models require a significant number of training samples, with a tendency towards unstable behavior, especially if the models are autoregressive. In this aim, we propose an adaptive methodology that combines traditional numerical solvers like finite elements, spectral methods, etc. with machine learning models. This methodology uses the traditional solver to generate training data for the ML model that will predict the flow dynamics. When the model uncertainty is sufficiently high the traditional solver is called again to generate new data.