Taishan Zhu¹, Jeffrey C. Grossman¹; ¹Massachusetts Institute of Technology

Strongly anharmonic solids have been widely sought, ranging from thermoelectrics to multiferroics and to shape-memory materials, but the effects of strong anharmonicity to their thermodynamic and transport properties are less understood. In this work, we extend our earlier crystal graph convolutional neural network framework, and explore an alternative Bayesian description for the vibrational modes and their transport in strongly anharmonic solids exhibiting soft modes. We demonstrate our theory on a one-dimensional toy lattice, including a probabilistic extension to the conventional phononic dispersion relationship, as well as its consequences to lattice conductivity, other thermal properties, and phase transition. This Bayesian framework also provides naturally an uncertainty quantifying scheme for these physical quantities. Our theory is then extended to two halide perovskites: all inorganic CsPbI3 and hybrid inorganic-organic MAPbI3, and our Bayesian predictions are compared with existing experiments of heat capacity and lattice conductivity.