Comparison of Active Metamodelling Techniques in Multidisciplinary Optimisation Frameworks

This research deals with the comparison of dynamic metamodels based on Radial Basis Functions and Artificial Neural Networks. The relevant framework is the robust optimal design in aeronautics under aeroacoustic objectives and constraints. This class of applications is constrained by the computational burden required by high–fidelity solvers to guarantee accurate solutions and the high number of function evaluations needed by the optimiser to converge. Consequently, the identification of efficient metamodelling techniques represents a crucial aspect for the designers. The use of metamodels can significantly reduce the number of high-fidelity evaluations, alleviating the overall computing costs. Accordingly, the engineering community has gradually switched from the design approach based only on direct simulations to the extensive use of metamodelling techniques. Recently, to make the metamodelling process even more efficient, function-adaptive strategies have been developed to improve the fitting capabilities. The dynamic properties of such approaches are mainly related to the self–tuning of the algorithmic parameters and the adaptive sampling of the domain. Here, dynamic metamodels based on Radial Basis Functions and Artificial Neural Networks are formalised and used to model the noise shielding properties of Blended Wing Body aircraft configurations. Special attention is paid to the definition of the metamodel uncertainty, which estimates the surrogate model goodness outside the known points. Both the formulations are demonstrated to correctly reproduce the phenomenon dynamics, although with substantial differences in the convergence properties.