Data-driven surrogates are the most popular replacement models utilized in many fields of engineering and science, including design of microwave and antenna structures. The primary practical issue is a curse of dimensionality, which limits the number of independent parameters that can be accounted for in the modeling process. Recently, a performance-driven modeling technique has been proposed where the constrained domain of the model is spanned by a set of reference designs optimized with respect to selected figures of interest. This approach allows for significant improvement of prediction power of the surrogates without the necessity of reducing the parameter ranges. Yet uniform allocation of the training data samples in the constrained domain remains a problem. Here, a novel design of experiments technique ensuring better sample uniformity is proposed. Our approach involves uniform sampling on the domain-spanning manifold and linear transformation of the remaining sample vector components onto orthogonal directions with respect to the manifold. Two antenna examples are provided to demonstrate the advantages of the technique, including application case studies (antenna optimization). KEYWORDS antenna design, constrained modeling, data-driven modeling, design of experiments, simulationbased design, uniform sampling 1 | INTRODUCTIONThe most versatile and ubiquitous antenna design tools nowadays are full-wave electromagnetic (EM) simulators. EM analysis permits reliable performance evaluation when executed at sufficient discretization level of the structure. EMdriven design closure (primarily, adjustment of geometry parameters) is mandatory yet challenging stage of the design process. The primary problem is a high cost of simulation, which may be acceptable for simple designs but not so much for complex structures described by a large number of parameters. In particular, numerous evaluations required by, eg, conventional optimization algorithms, 1-5 may be impractical. The problem is even more pronounced for tasks involving massive simulations such as statistical analysis 6 or tolerance-aware design. 7,8 The most common work-around is an interactive design based on parameter sweeping; however, this approach has serious limitations: it fails to yield optimum designs, cannot handle design constraints, or cannot account for parameter interactions, to name just a few.