Radial basis function simulation and metamodelling of surface roughness in centreless grinding

Purpose: The purpose of this study was to investigate the efficiency of artificial neural networks and the related metamodels to simulate and identify complex centreless grinding process. Design/methodology/approach: The modeling is founded on the system approach, which is efficiently dealing with the complexity of the grinding process. The unknown process transfer function is identified via artificial neural network that requires fewer assumptions and less precise information about the process modeled than other conventional modeling techniques. The developed metamodel is a response surface (polynomialfit) of the simulated process that is achieved by the computer model. Findings: The metamodel quality is strongly related to the prediction accuracy of the underlying simulation model. The generalization capability of an artificial neural network is sensitive to the training samples (design of experiments). The predictive ability of a metamodel is comparable to the accuracy of the response surface regression model. Research limitations/implications: Improved simulation model and application of unconventional metamodels (Gaussian process regression) will significantly improve the presented preliminary results. Originality/value: Metamodelling of computer experiments is an expansion of response surface methodology and the classical designs of experiments and represents a new paradigm in empirical modelling of machining operations.