Domain Segmentation Based On Uncertainty In The Surrogate (DSUS)

This paper develops a novel approach to characterize the uncertainty in the accuracy of surrogate models. This technique segregates the design domain based on the level of cross-validation errors; the overall framework is called Domain Segmentation based on Uncertainty in the Surrogate (DSUS). The estimated errors are classified into physically meaningful classes based on the user's understanding of the system and/or the accuracy requirements for the concerned system analysis. In each class, the distribution of the cross-validation errors is estimated to represent the uncertainty in the surrogate. Support Vector Machine (SVM) is implemented to determine the boundaries between error classes, and to classify any new design (point) into a meaningful class. The DSUS framework is illustrated using two different surrogate modeling methods: (i) the Kriging method, and (ii) the Adaptive Hybrid Functions (AHF). We apply the DSUS framework to a series of standard problems and engineering problems. The results show that the DSUS framework can successfully classify the design domain and quantify the uncertainty (prediction errors) in surrogates. More than 90% of the test points could be accurately classified into its error class. In real life engineering design, where we use predictive models with different levels of fidelity, the knowledge of the level of error and uncertainty at any location inside the design space is uniquely helpful.