A Comparison of Polynomial Response Surfaces and Gaussian Processes as Metamodels for Uncertainty Analysis with Long-Running Computer Codes

摘要

We discuss some of the findings of a recently completed extensive review of the literature concerning methods for approximating the input-output behavior of complex computer codes (i.e., metamodeling). This review has been motivated by a need in the nuclear safety community to perform uncertainty analysis (UA) for long-running (e.g., several hours or days per simulation), mechanistic computer models, a task that has often proven prohibitively expensive due to the need to perform hundreds or thousands of simulations. We focus on the use of polynomial response surfaces (RS) and Gaussian processes (GP) as metamodels. RSs, though simple, efficient, and often quite useful, have been criticized by several authors who have raised numerous objections to their use as metamodels for deterministic codes, and we summarize these objections. GPs offer several advantages over RSs, particularly with regards to these criticisms, but do so at the expense of some simplicity and efficiency. Finally, we present the results of a case study involving the failure probability estimation of a passive nuclear safety system. The results of this case study suggest that GPs are superior when data are limited, but that both RSs and GPs perform comparably when many data are available.