Flow passing a heated square cylinder is investigated using a hybrid LES-RANS approach on unstructured grids at a moderate Reynolds number of 22,050.The implicit sub-grid scale model is apply to LES, and two turbulence models, the Spalart-Allmaras model and the SST k-omega model, are employed for near-wall RANS. Combined with LES, both models present good predictions of the time-and phase-averaged velocity profiles on a 4-million-cell grid. Results of the LES-SST approach agree better with the experimental data especially at locations close to the cylinder surface and this leads to improved surface convective heat transfer compared to LES-SA. Grid convergence study shows that grid resolution in the near-wall region and on the cylinder surfaces is important in resolving the unsteady convective heat transfer. Results of velocity field and surface heat transfer from the fine grid with 8 million cells compare favourably with the experimental data and show significant improvement over that of the medium and coarse grids. Analysis of turbulent statistics is performed by means of energy spectra and anisotropy invariants of the Reynolds stress tensor. Proper orthogonal decomposition (POD) is used to identify the vortex shedding phases. It is shown that the POD based phase-averaging produces more accurate velocity profiles than the conventional pressure-signal based method. (C) 2016 Elsevier Ltd. All rights reserved.