Comparison of ANN and DoE for the prediction of laser-machined micro-channel dimensions

Abstract

This paper presents four models developed for the prediction of the width and depth dimensions of CO 2 laser-formed micro-channels in glass. A 3 3 statistical design of experiments (DoE) model was built and conducted with the power ( P), pulse repetition frequency (PRF), and traverse speed ( U) of the laser machine as the selected parameters for investigation. Three feed-forward, back-propagation artificial neural networks (ANNs) models were also generated. These ANN models were varied to investigate the influence of variations in the number and the selection of training data. Model A was constructed with 24 data randomly selected from the experimental results, leaving three data points for model testing; Model B was constructed with the eight corner points of the experimental data space, and seven other randomly selected data, leaving 12 data points for testing; and Model C was constructed with 15 randomly selected data leaving 12 data points for testing. These models were developed separately for both micro-channel width and depth prediction. These ANN models were constructed in LabVIEW coding. The performance of these ANN models and the DoE model were compared. When compared with the actual results two of the ANN models showed greater average percentage error than the DoE model. The other ANN model showed an improved predictive capability that was approximately twice as good as that provided from the DoE model.