This work implements autoregressive machine learning models to learn unsteady flow dynamics around 2D geometries at different Reynolds numbers. Leveraging convolutional, recurrent and/or transformer architectures, the models are trained on CFD snapshot data to learn temporal flow evolution. This approach enables rapid prediction of flow properties given an initial state, especially in unsteady flows where fine spatial and temporal resolution is needed. The pipeline supports 2D unstructured meshes and scalar input parameters. Preliminary results on the unsteady flow around a 2D cylinder demonstrate accurate short-term flow reconstruction and indicate potential for generalization across geometries, contributing to the development of data-driven models for unsteady flow prediction in computational fluid dynamics. Modeling is studied exploring several techniques for the spatiotemporal representation of the data and loss function definition in order to to learn it.