Neural network fitting and prediction of polarisability tensors for aromatic hydrocarbons

A standard back‐propagation neural network is used to correct input semi‐empirical molecular orbital calculations of polarisability tensors to fit experimental data for aromatic hydrocarbons. The method readily yields the correct component normal to the molecular plane but is restricted by a small training set. The network is also used to predict polarisability components for structures input as the pattern of rings fused to a central benzene ring. Semi‐quantitative predictions are obtained depending on the size and method of presentation of the training set.