Predicting Stable Electron-Exact Clathrate Superstructures via Convolutional Nerual Network

Clathrate structures consist of nanometer-size polyhedral cages encapsulating guest atoms or molecules that are not directionally bonded to the framework. Compounds in this class of crystal structure have already demonstrated many unique properties, such as ultra-low thermal conductivity and superconductivity. Enforcing electron-balanced compositions, with an average of four electrons per framework site, results in a semiconducting electronic structure and offers utility in electronic devices. To date, only a handful of stable electroneutral tetrel-free clathrates are known, most as ordered superstructures up to eight times larger in volume than the type-I clathrate unit. In this work, new clathrate compounds are predicted computationally and experimentally verified through synthesis. The group III-V A8T27Pn19 clathrate family (A={Na, K, Cs, Rb}, T={Al, Ga, In}, Pn={P, As, Sb, Bi}) is studied comprehensively with density functional theory (DFT). A crystal graph convolutional neural network (CGCNN) model is trained on the computed clathrates and roughly 500 compounds of similar composition. The model is refined by the results of experimental synthesis and employed to discover other electron-exact superstructures composed from group III-V or II-VI elements.