In the framework of nondestructive eddy-current testing, an inversion method is proposed. It relies on a metamodel-based optimization method: a fast optimization by means of a particle swarm algorithm is performed wherein the exact forward computations are replaced by an appropriate metamodel. The latter is obtained by radial basis function interpolation over a database previously generated by sequential design. Both synthetically generated and laboratory-controlled experimental data are used to illustrate its performances. Comparisons to inversions by support vector machines, now common for defect characterization, are carried out. The proposed method appears to be an useful tool for decision analysis, highlighting the likeliest results.