Prediction of the Impact Sensitivity by Neural Networks

Abstract

A method for optimizing the prediction of impact sensitivity of explosive molecules by neural networks is presented. The database consists of 204 molecules of known sensitivity, containing C, H, N, and O and belonging to several chemical families. Pertinent molecular descriptors are selected by a preliminary evolutionary multiple linear regression treatment, and the effects of the network's topology and the extent of the training are examined and optimized. The predictions are satisfactory with a correlation coefficient R = 0.94 obtained through cross-validation. The neural networks approach proves more accurate than linear methods and more general than all previously used methods.