Application of artificial neural network and global optimization techniques for high throughput modeling of the crystal structure of stannites and kesterites

This study aims to apply artificial neural networks for the prediction of the lattice parameters of materials with stannite‐ and kesterite‐type structure, and to compare the results of predictions with that obtained in the calculations exploiting the density functional theory. Crystallographic data for 49 compounds with stannite‐type structure and for four compounds with the kesterite‐type structure are found and, based on it, crystal structures are calculated using the density functional theory (DFT) method in a two‐step relaxation procedure for all compounds. An multilayer Perceptron is constructed, which then is trained on gathered crystallographic data. Values predicted by a neural network (lattice parameters) are compared with experimental data and with results of DFT calculations. Moreover, a global optimization method (the Uspex code) is used to find potentially novel crystal structures for investigated chemical compositions. The results are discussed in the term of advantages and disadvantages of each method. DFT optimization, Uspex algorithm, and artificial neural network are used for the crystal structure prediction. Artificial neural network outperforms DFT optimization in lattice parameter predictions. Moreover, permutation importance algorithm reveals the influence of each input variable. It turns out that for stannite structure, more important are electron than geometrical properties. The Uspex algorithm is capable of prediction of experimentally observed stannite and kesterite structures, while suggesting the existence of novel structures.