Neural network versus Kriging, surrogate models for LCM process optimisation

Abstract

In Liquid Composite Moulding (LCM) processes an optimal combination of a variety of manufacturing design variables must be chosen in order to minimize cycle time while keeping equipment, layout and running costs low. Such black-box function optimization can be achieved by integrating the process simulation algorithm with a metaheuristic, such as a genetic algorithm (GA). However, the large number of function evaluations required by a GA combined with the computational expense of the simulations often makes this approach unaffordable. This issue is further emphasized when the filling and curing phases are coupled and an iterative optimization strategy becomes necessary. The use of a surrogate model, as a substitute to the expensive simulation algorithm, is a common technique for reducing the run-time of an optimization algorithm. Choosing such a model that suitably duplicates all the features and trends of the objective functions over the design space, is an important task. In this paper we compare two popular surrogate models, namely the artificial neural network (specifically the Cascade-Correlation Learning Architecture Neural Network) and kriging, and discuss their performance in terms of prediction accuracy and the run-time of the resultant optimization algorithm.