In naval architecture many functional surfaces such as hulls, foils, rudders and propellers feature complex shapes with compound curvature. In simulation-driven design these shapes are subject to variation. Since many variants are investigated during design space exploration and for optimization campaigns robust models are needed that yield very low failure rates, preferably zero failure, meaning 100% robustness. In addition, often quite a few constraints have to be met which calls for nested optimizations and for comprehensive checks before launching any expensive simulation. The time spent and the workforce bound to create, repair and review shapes interactively are too large, preventing traditional Computer Aided Design systems with a history-based modeling approach to be effective. Rather, a dedicated approach is needed that allows combining a range of modeling techniques and readily incorporating constraints. Using the data that are automatically generated, for instance by means of a Design-of-Experiment (DoE), machine learning (ML) models can be trained. The presentation will cover techniques of defining robust parametric models for simulation- and data-driven design as implemented in CAESES, discussing examples from sports (hull for 37th America's Cup), recreation (tip-rake propeller for a battery-powered super-light planing boat) and industry (hull form of an electric catamaran). Each of the three examples for robust parametric models will be used to showcase important aspects: The AC75 model is highly-complex and needs to observe a large number of constraints, the propeller model is fully-parametric and incorporates all aspects of the geometry directly while the model for the electric boat is hybrid with fully- and partially-parametric modeling components. Each example will be discussed in the context of its practical application and the results achieved at full-scale.