Optimisation of radial basis and backpropagation neural networks for modelling auto-ignition temperature by quantitative-structure property relationships

Quantitative structure-property relationship (QSPR) modelling was successfully employed for the prediction of auto-ignition temperatures (AIT) via neural networks using the structural and physicochemical properties of 233 organic compounds. Two types of feed forward neural modelling networks, radial basis function (RBF) and backpropagation function (BPF), were applied and compared based on their ease of model building, prediction accuracy, and network simplicity. Optimisation of the spread parameter and neurons in the hidden layer for RBF networks by 32 factorial response surface methodology enabled a rapid convergence to the best network configuration for modelling AIT. Because of the random initialisation of weights in BPF networks, traditional factorial optimisation could not be utilised, however a 32 factorial design for training time and number of neurons in the hidden layer was applied in an intensive replicative fashion for each design point to achieve a model with good predictive ability. Using a training set of 85 compounds with 13 different functional groups, AIT values for the rest of the validation set of 148 compounds were successfully predicted to within the laboratory experimental error of ± 30°C. The RBF network was found to be superior to the BPF network for this application in terms of speed and ease of optimisation, and the RBF network was also found to be less susceptible to the presence of local minima. Comparison of neural modelling with multiple linear regression and partial least squares shows that better prediction accuracy can be obtained by neural networks.