Neural networks applied to regioselectivity of radical additions

A multilayer neural network has been successfully applied for predicting the regioselectivity in radical additions. A three-layered feed-forward network with the back-propagation algorithm was employed. The data for the input layer were the quantitative steric descriptors of substituents on alkenes and of radicals, the frontier molecular orbital energies and the charges of olefinic carbon atoms. The output data were the logarithm of experimental product ratios. The obtained results have pointed out that not only the major products but also the ratios of products have been correctly predicted by using neural network architecture