To ensure the high reliability of a composite structure, the actual mechanical properties of a material must be accurately predicted. Here the problem of identification of elastic properties of composite structural elements using vibration tests is considered. The metamodels built on the basis of FEM computer experiments and physical measurements of eigenfrequencies are used for the identification Two methods are compared. The first, a well known method, is the minimization of the discrepancy between calculated and measured frequencies. Eigenfrequencies are calculated by a finite element model using a numerical experiment – a set of trial values for the unknown material parameters. The numerical frequencies are compared with the measured frequencies, and material properties are found by minimizing the relative discrepancy. For this method the analysis of errors, caused by manufacturing errors, measurement errors and the lack of coincidence of FEM and physical model is proposed. The second method is the direct building of the inverse metamodel x=F(f), where x – vector of material properties and f – vector of eigenfrequencies. In this case the inputs are highly correlated, but the local approximation methods give the possibility to find good approximations. The significance analysis can help to choose the best subset of eigenmodes for the first method. As example, three elastic moduli of four specimens which are cut out from a large curved stiffened composite panel were identified. The results of both methods are very close. The identification results obtained from the vibration tests of the small panel slightly differ from the typical material properties of CFRP composites, which can be explained by the fact that single ply thickness, material density, and layer angles of the real structure are different from the nominal values. The sensitivity analysis shows that for the adequate determination of material properties with errors around 5-10 percent, all parameters, not only frequencies, must be measured with split-hair accuracy – error not greater than 0.5 – 1 percent. To avoid the bias error, the effect of external and internal energy dissipation should be taken into account.