Multidisciplinary optimization systems rely increasingly on parametric CAD engines to supply the geometries required by their analysis components. Such parametric geometry models usually result from an uneasy compromise between high flexibility, that is, the ability to morph into a wide variety of topologies and shapes, and robustness, the ability to produce feasible, sensible topologies and shapes throughout most of the design space. It is argued that a possible means of achieving both objectives is via a supervised learning system attached to the CAD model. It is shown that such a model can capture some of the engineering and geometrical judgment of the designer and can thereafter be used to repair design variable sets that lead to infeasible CAD models.