Vol. 45, Issue 4, pp. 545-557

Abstract

Artificial neural networks have been shown to be able to approximate any continuous nonlinear functions and have been used to build data based empirical models for nonlinear processes. This work studies primarily the performance of neural networks as a tool for predicting the emission spectra of fluorescent materials from their absorbance, and further, tends to the determination of the optimal topology of the neural network for this purpose. In order to do this, spectral data were initially analyzed by a principal component analysis technique. The first four principal components were used as input nodes of neural networks with various training algorithms – namely cascade- and feed-forward algorithms – and also, various numbers of hidden layers and nodes. The obtained results indicate that the RMS error in a testing data set decreased with increasing the number of neurons and the minimal network architecture for a data prediction problem consists of two hidden layers, respectively with 9 and 1 nodes for both neural networks. Additionally, a better performance was obtained with the cascade-forward neural network, especially in a small number of nodes. The obtained results indicate that the neural networks can be used to provide a relationship between the absorbance as an input and the emission as a target.