“…From simulation data, a model y = f (x) is constructed, where y is typically a performance metric, x includes design, process, or environmental variables, and f is an approximation of the SPICE mapping. Models used include linear models [8], [9], [25], posynomials [10]- [12], polynomials [13], [14], [25], splines [15], [25], neural networks [16], [17], [25], boosted neural networks [18], [25], support vector machines [19]- [21], [25], latent variable regression (LVR) [22], [23], kriging [24], [25], and stochastic gradient boosting [26]. However, such models either follow an overly restrictive functional template which limits their applicability, or they are opaque and thus provide no insight to the designer.…”