Predictive modeling is an important tool in engineering design and optimization. Designers can develop a predictive model to replace a computationally-intensive physics-based model (a practice referred to as meta-modeling or response-surface modeling) or to model systems based on empirically-obtained data. However, such models typically have a limited domain of validity—that is, only certain combinations of model inputs yield predictions that are trustable. Consequently, designers must take care to bound the search space of optimization algorithms that otherwise would be unable to distinguish between valid and invalid predictions. Prior research has found that the valid input domain of a model can be shaped irregularly and difficult to model using simple bounds on input variables. The Support Vector Domain Description (SVDD) method was shown to be an effective approach for modeling such boundaries. However, the method used previously for generating the domain description is slow and scales poorly as the size of the training data set grows. This paper describes a new incremental method for generating a SVDD using a point-by-point comparison in place of considering all data points at once. This method is observed to be over 1000 times faster than the original method. This makes the overall approach attractive on problems of practical scale. We describe the new method, explore its characteristics, and demonstrate it on a design example for the selection of component concepts for a commercial power generation plant.

This content is only available via PDF.
You do not currently have access to this content.