A methodfor optimizing the prediction ofimpact sensitivity ofexplosive molecules hv neural networks is presented. The database we used consisted 0/272 molecules containing C,H,N,0 ofknown sensitivity and belonging to several chemicalfamilies. Pertinent molecular desriptors were selected by a preliminary multilinear treatment. The effects of the network's topology, the extent of the training, the choice of descriptors were examined and optimized. The predictions are satisfactory with a correlation coefficient of 0.94 obtained through cross-validation. Moreover 95% of compounds are correctiy classißed in a 3sensitivity scale and the remaining 5% are classißed äs ambiguous which is very encouragingfor a real worid implementation. The neural networks approach proves more accurate and more general than previous methods.

No abstract