Convex metamodels for reservoir optimization

Abstract

A metamodel, or surrogate model, is a simplified model of an actual model of a system. Metamodels are particularly important when the true model is expensive to run, as they typically map a set of inputs to an expected output through regressions, which are, in general, quick to evaluate. It is possible to consider local metamodels, where the validity of the results is expected only in the vicinity of some point, or global metamodels, which define a relationship that is valid across all feasible input values. In a variety of applications, many regression methods have been used to fit global metamodels, including radial basis functions, gaussian processes, stochastic kriging, and neural networks. Metamodel optimization is used for inferring optimal decisions from observational data generated by a black-box simulator. We propose hereto evaluate the quality of the main convex metamodels in use now in the UNISIM research group. (AU)