This paper focuses on the use of transfer learning for training two different Machine Learning algorithms for damage detection and localization, respectively. It is well known that the learning phase benefits from an accurate representation of the intact structure, a realistic noise contamination of the synthetic outputs produced via the virtual experiments and, if needed, the parametric inclusion of damage scenarios. The selected modal features can be advantageously based on those involved in physics-based damage identification to achieve a higher accuracy. In the comparison with more basic features, such as modal shapes and curvatures, less attention has been given to the effects of modeling error. The above principles and analyses are applied to the identification of damage in a rectangular metallic plate for which measured modal data is available from a dedicated experimental campaign in CNR-INM laboratory using accelerometers. The damage is modeled as a stiffness reduction over a small area, being also representative of material thinning due to corrosion. To have meaningful training database, the modal convergence of the FE model to the real structure is guaranteed by a structural optimization process. Experimentally identified noise, representative of real-life applications, is then added to the FE results before algorithm training. Damage existence and position are determined by a Novelty Detection approach and a Regression Neural Network, respectively. It will be shown that basic features always provide less accurate damage estimation in presence of modeling errors on boundary conditions with respect to more structured ones.