An efficient adaptive database sampling strategy with applications to eddy current signals

Abstract

Computer simulations are widely used in engineering domains to model complex scenarios and extract meaningful information or improve the understanding of a given problem. Common purposes of simulation studies are inversion, optimization, sensitivity analysis and evaluation of performance. In such contexts, it is often convenient to replace the time consuming forward solver by a metamodel acting as a fast and accurate substitute in a restricted range of input parameters. Focused on applications in the field of Electromagnetic-Non Destructive Testing (E-NDT), this paper proposes an approach to design robust metamodels, based on adaptive databases of simulation results in order to ensure their accuracy. They can then be used as real-time emulators of the physical model and considerably speed up time consuming studies like estimation of probability of detection, defect characterization or sensitivity analysis. The database and metamodel generation problem is first addressed with a meshless approach based on Augmented Radial Basis Function (A-RBF) algorithm. Then, its performance is compared with that of a more standard approach exploiting a n-dimensional Delaunay mesh. Both approaches rely on an adaptive generation technique known in the literature as Output Space Filling (OSF). Performance in terms of computational time and results accuracy of both methods are finally evaluated and compared in the case of a specific application: the simulation of Eddy Current Testing (ECT) inspection problems.