Neural network for prediction of 13C NMR chemical shifts of fullerene C60 mono-adducts

Real-valued models based on deep artificial neural networks were proposed to predict C-13 NMR chemical shifts of fullerene C-60 core carbon atoms for computer-aided structure elucidation of complex fullerene C-60 mono-adducts. We showed that parametric rectified linear units could be successfully used as activation functions in hidden layers of artificial neural networks for decision of complex physical-chemical tasks. A total of 400 artificial neural networks were trained and tested in order to reveal the best-fitted models. The best prediction accuracy of real-valued models was achieved with MAEP=1.83ppm/RMSEP=2.60ppm using artificial neural network model which has 110 and 120 hidden units, respectively, with parametric rectified linear unit as activation function. A complex set of atomic descriptors for fullerene core carbons is suggested based on modern approaches. Real-valued C-13 NMR shifts predictor based on neural network is put forward for complex fullerene C-60 mono-adducts. Parametric rectified linear unit activation function is shown as suitable for C-13 NMR prediction models for complex fullerene C-60 derivatives.