Evolutionary aerodynamic shape optimization through the RBF4AERO platform

This paper demonstrates a way of solving industrial aerodynamic shape optimization problems using the RBF4AERO platform, developed in the framework of the EU–funded RBF4AERO project. The platform combines optimization algorithms (stochastic and gradient– based ones), a mesh morphing tool based on Radial Basis Functions (RBFs) and various evaluation tools (CFD, CSD, etc). In this paper, the Evolutionary Algorithms (EAs) based tool assisted by metamodels trained on a sampling performed during the Design of Experiment phase is used along with a CFD evaluation tool. The use of Response Surface Models (RSM) significantly reduces the number of CFD runs required to reach the optimal solution(s), while the use of RBF–based mesh morphing avoids re–meshing prior to each CFD–based evaluation. In optimization problems, the platform starts by selecting a number of individuals to undergo CFD– based evaluations. The latter constitute the training patterns for the RSM which is, then, used as a low-cost evaluation tool within the EA–based optimization. " Optimal " solutions found by the EA–based search exclusively based on the trained metamodel are then re-evaluated by means of the CFD tool. The database of already evaluated individuals is updated, the RSM is re–trained and the EA–based optimization is repeated. The optimization terminates when convergence criteria related to the RSM prediction accuracy are met or the computational budget is exhausted. The optimization of an ultra-light aircraft and that of the DrivAer car model for minimum drag are showcased. In all cases presented, the simpleFoam solver of OpenFOAM is used to evaluate candidate solutions.This paper demonstrates a way of solving industrial aerodynamic shape optimization problems using the RBF4AERO platform, developed in the framework of the EU-funded RBF4AERO project. The platform combines optimization algorithms (stochastic and gradient-based ones), a mesh morphing tool based on Radial Basis Functions (RBFs) and various evaluation tools (CFD, CSD, etc). In this paper, the Evolutionary Algorithms (EAs) based tool assisted by metamodels trained on a sampling performed during the Design of Experiment phase is used along with a CFD evaluation tool. The use of Response Surface Models (RSM) significantly reduces the number of CFD runs required to reach the optimal solution(s), while the use of RBF-based mesh morphing avoids re-meshing prior to each CFD-based evaluation. In optimization problems, the platform starts by selecting a number of individuals to undergo CFD-based evaluations. The latter constitute the training patterns for the RSM which is, then, used as a low-cost evaluation tool within the EA-based optimization. "Optimal" solutions found by the EA-based search exclusively based on the trained metamodel are then re-evaluated by means of the CFD tool. The database of already evaluated individuals is updated, the RSM is re-trained and the EA-based optimization is repeated. The optimization terminates when convergence criteria related to the RSM prediction accuracy are met or the computational budget is exhausted. The optimization of an ultra-light aircraft and that of the DrivAer car model for minimum drag are showcased. In all cases presented, the simpleFoam solver of OpenFOAM is used to evaluate candidate solutions