Turbulence modeling is a challenging topic in fluid dynamics, with implication for numerous engineering applications. Recent studies have proposed new path of investigation with usage of Machine Learning (ML) approaches. Such tools permit to exploit available high-fidelity data to infer new turbulence closures or to augment existing ones. However, the required data for ML training, which is generally obtained from direct numerical simulation (DNS), may be expansive to produce and restricted to specific geometries and relatively low Reynolds (Re) numbers [1]. On the other hand, experimental data may be available for flows with various set-ups and a large range of Re numbers but may be affected by noise and limited in terms of spatio-temporal resolution. Dealing with sparse data for ML training is a difficult task [2], and it is a topic still relatively unexplored in the context of Computational Fluid Dynamics (CFD). Furthermore, ML applications for unsteady flows are still at embryonic stage. In this study, we present a method to train a ML-based turbulence model for an Unsteady Reynolds Averaged Navier-Stokes (URANS) simulation from noisy observations that are sparse in both space and time (Fig1). The methodology builds upon a Bayesian approach [3], relying on ensemble data assimilation (DA) methods to reconstruct full fields from observations (Fig2). Subsequently, a ML-based model is trained to augment the baseline URANS simulation (Fig3). The test case chosen for this study is the unsteady flow around a circular cylinder for Re=3900.