Vibrational properties of solids : a machine learning approach

This work presents a novel approach based on a neural network and transfer learning for the prediction of thermodynamic properties of crystalline solids. It has the unique property of taking advantage of information stored across different properties, in order to increase overall accuracy. Indeed, vibrational properties of solids are key for material design and discovery. For instance, the vibrational entropy can easily alter the structure stability. Although ab initio high-throughput calculations allow for the screening of properties of thousands of materials, fast characterization and prediction of thermodynamic properties of materials is still missing due to the inherent computational cost involved in obtaining second-order derivatives in energy. Thanks to the presented thermal transfer neural network, a fast on-the-fly prediction of the vibrational entropy, energy, specific heat and Helmoltz free energy at various temperatures are possible and this for any crystal structure with a mean absolute error three times as low than previous models. Furthermore, this dissertation proposes an interpretable feature selection algorithm based on a dynamic relevance-redundancy criterion. Finally, trends are found between common material features and important causes to the vibrational entropy are stated, such as the radial environment of a structure's sites.