Multi-channel convolutional neural networks for materials properties prediction

Deep convolution neural networks (ConvNets) have been recently used to predict the enthalpy of formation and the prediction errors are within DFT precision. Here we show that a multi-channel input for the ConvNets improves the prediction accuracy, and the accuracy can be further improved by decomposing the input signals into high/low frequencies. We trained ConvNets using the periodic table representation on a DFT formation enthalpy dataset of 10,590 elpasolite compounds. The mean absolute error (MAE) reaches 50 meV/atom, which is half value of the MAE of a ConvNet using single input channel. The dependence of MAE on each element was also analyzed. Our work demonstrates the importance of input data preprocessing for ConvNets prediction accuracy in material informatics tasks.