<p>Numerical experiments of targeted grid refinement are reported for stress-based wall-modeled large-eddy simulation (WMLES) of turbulent smooth-body separation. The flow over a Gaussian bump at a length-based (height-based) Reynolds number of 2 million (0.17 million) is simulated using an unstructured polyhedral solver on isotropic grids, and compared to high-fidelity Direct Numerical Simulation (DNS) data. The baseline coarse grid with 16 points per boundary-layer thickness in the upstream region does not capture smooth-body separation, while the baseline fine grid with double the resolution over half the boundary layer accurately captures the separated flow region indicating the importance of near-wall refinement. Refining the regions in the vicinity of the apex of the bump where the pressure gradient effects are expected to be dominant gave similar predictions to the baseline fine grid. Past DNS studies have noted the presence of an internal layer that begins to develop in the favorable pressure gradient region upstream of the bump apex, where its thickness is under 10% of the local boundary-layer thickness. However, the importance of sufficiently resolving the internal layer in predicting the flow separation accurately is unclear. Near-wall refinement just upstream of the apex targeted at the developing internal layer did not produce any flow separation, while near-wall refinement downstream of the apex in the vicinity of separation produced reasonable predictions, indicating that the latter is more critical for WMLES. While the present simulations contain uncertainties/errors due to the choice of grid topology, subgrid and wall model, these targeted grid refinement results provide useful insights to design optimal grids to capture turbulent smooth-body flow separation.</p>