Reducing the cost of computational fluid dynamics optimization using multi layer perceptrons

Abstract

The paper presents a method for reducing the cost of Computational Fluid Dynamics optimization by using a neural network to fill-in the design space. The method trains a network to approximate the aero-or hydrodynamic performance of vehicles with the Cascade Correlation algorithm. This network is coupled with a Genetic algorithm to optimize the hydrodynamic performance of the configuration. I.U>, which, in CFD, is the aero-or hydrodynamic performance for a given configuration.A general optimization process is illustrated in Fig. 1. An initial set of design variables, which hight represent the configuration designed by experienced engineers, is supplied to the optimizer. Then, for this design, the objective function, f; is evaluated and the constraints, gi, are analyzed to check whether they are violated or not. If the optimum is not reached, these values are fed back to the optimizer that modifies the D.V.'s. The process is repeated until convergence.

Adaptive-surrogate based on a neuro-fuzzy network and granular computing

“…Techniques based on universal approximators, whose quality depends on the training data, are based on Artificial Neural Networks (ANN) [9,11,14,17,24,25]. In [15], selection of centers in the RBF surrogate model, is done in an unsupervised manner with Learning Vector Quantization (LVQ) and Self-Organizing Maps, this formulation tackles the problem of good generalization, that represents an estimation of objective functions for new individuals.…”