[Prediction of retention parameters of explosives by artificial neural network]

The quantitative relationship between the retention parameters and the structure of explosives is discussed. Molecular topological indices are used to represent the structure. Based on the back-propagation algorithm, a quantitative model was established after a training process of a train-set containing 20 explosives being completed. The Sigmoid function was chosen as the transmit function. The retention parameters (log kappa w and S) acted as output vectors, while molecular connecting indices (0 chi, 1 chi, 2 chi) and edge adjacent indices(epsilon) acted as input vectors. The input vectors were normalized in the range of -3-3 and the output vectors were normalized in the range of 0-1. The accuracy of network was 0.5 and the beginning value of studying pace (eta) was 0.2, the momentum factor (alpha) was 0.5. The results showed that the yield model reflected the relationship between the structure and retention index of compounds, and had high accuracy. Most of the relative errors were below +/- 5%.