Data-driven compression of first-principles phonon-phonon interactions

Phonon-phonon (ph-ph) interactions originating from the anharmonicity of interatomic potentials play a crucial role in various properties of crystals such as thermal conductivity and thermoelectricity. First-principles calculations can provide accurate 3rd- and 4th-order interatomic force constants (IFCs) governing ph-ph interactions. However, these IFCs are large tensors containing millions to billions of entries, and constitute a major bottleneck in phonon transport and nonequilibrium lattice dynamics calculations.



In this talk, we present an efficient approach to manipulate 3rd-order IFCs using a widely used compression technique, singular value decomposition (SVD). We demonstrate that keeping only 1-2% of the singular values allows us to obtain accurate ph-ph interactions on dense momentum grids with a significant speed-up compared to state-of-the-art calculations. This allows us to more efficiently compute the thermal conductivity and related properties, as we show for several materials including silicon, MgO, and TiNiSn. Our compression technique enables significantly faster first-principles calculations of phonon lifetimes, thermal conductivity, ultrafast phonon dynamics, and other properties governed by ph-ph interactions.