Ryotaro Okabe1,Zhantao Chen1,Mingda Li1
Massachusetts Institute of Technology1
Ryotaro Okabe1,Zhantao Chen1,Mingda Li1
Massachusetts Institute of Technology1
The phonon<sup>1</sup> of a material describes the dynamics of its constituent atoms in the harmonic approximation, in the framework of the theory of lattice vibrations. The details of the lattice dynamics are of key importance. Especially, phonon at the Gamma point of the reciprocal space gives us important properties such as dielectric constant and the stability of materials.<br/> <br/>In this work, we present a machine-learning model that directly predicts phonon band structure only with simple input like atomic coordinates. At the beginning of the machine learning model, the input feature vectors are passed to an embedding layer for dimensionality reduction. In the next part, the Euclidean neural network (E(3)NN)<sup>2</sup> is then applied to the resulting hidden state and consists of alternating convolution and gated block operations. Here, a convolutional layer operates on the radial distance vectors between atoms in a neighborhood up to a radial cutoff. E(3)NN has made it possible to include 3D rotation, inversion, and translation symmetry in the atomic structures so that high accuracy is reached without data augmentation.<br/> <br/>The method we present this time will give us insights about the materials’ dynamics. In addition, it is expected that we can expand the tool to predict properties that are difficult to predict, such as the magnetic structures<sup>3</sup>.<br/> <br/>Reference<br/>[1] Petretto et al., Sci Data 5, <b>2018</b>, 180065<br/>[2] Smidt et al., <i>Phys. Rev. Res.</i> <b>2021</b>, <i>3</i> (1), L012002.<br/>[3] Rodríguez-Carvajal, et al., <i>C. R. Physique</i> <b>2019</b>, <i>20</i>, 770–802.