Adaptive Kriging Metamodels for Expensive-to-Run Electromagnetic Simulations
Résumé
Metamodeling is getting more and more widespread in the domain of electromagnetics to replace the often computationally expensive, time-consuming numerical simulations. This paper presents a novel method of metamodeling based on the kriging interpolation. The input-output function (realized by a numerical simulator) is interpolated using a set of output observations at certain input values (“samples”). The quality of the yielded metamodel strongly depends on the choice of these samples. An algorithm is presented to select samples sequentially by optimizing a given criterion related to the optimality of the specific metamodel being built. This adaptive strategy is illustrated and compared to a classical approach via numerical examples drawn from electromagnetic nondestructive evaluation. The presented examples demonstrate that very precise metamodels can be achieved in the proposed way.