Machine Learning for Phonon Thermal Transport

Abstract

Efficient generation, transport, conversion, and storage of energy are essential to support our modern society and combat global climate change. As one of the major energy carriers, phonons play an indispensable role in various energy-related processes. The past decades witnessed continuous development in phonon measurements and computation techniques. However, at least two significant challenges impede our further exploration of phonon thermal transport. The first challenge is the difficulty of acquiring certain phonon properties in a streamlined way. Key quantities like density-of-states (DOS) are nontrivial to compute with high computational costs and measure with complicated experimental setups. Second, there is a lack of techniques to detect phonon frequency-based information. The phonon frequency-based information, such as relaxation time and interfacial transmittance, contains rich microscopic insight and governs measurables like thermal conductivity and interfacial thermal conductance. However, it is generally beyond the reach of existing measurement techniques. This thesis demonstrates how machine learning can address the challenges by 1) predicting phonon DOS from simple information of atomic structures and symmetry-aware neural networks and 2) extracting frequency-based phonon information from time-resolved diffraction measurements with scientific machine learning.

First, we present the direct prediction of phonon DOS using only atomic species and positions as input. We apply the symmetry-aware Euclidean neural networks (E(3)NN) to preserve crystallographic symmetries and overcome data scarcity. The predictive model reproduces essential features of phonon DOS and generalizes to materials with atom types absent from the training set. We further exemplify our method’s potential by predicting alloy systems without any additional computational cost and filtering high phononic specific heat materials out of more than 4,000 candidate materials within one hour. Second, we reveal microscopic phonon transport in heterostructures with a machine learning augmented experiment framework. By taking advantage of the ultrafast electron diffraction (UED) with the dual temporal and reciprocal-space resolution, we employ advanced scientific machine learning to recover the frequency-dependent interfacial transmittance with possible extension to the relaxation time of each layer. We demonstrate its capability by analyzing thin heterostructures beyond conventional experimental methods and reconstructing unprecedented details of real-space, real-time, frequency-resolved phonon dynamics across an interface.

While the presented topics are closely related to phonon systems, the proposed methods in this thesis can be readily transferred to predict other continuous properties and learn the previously “unmeasurable” properties from experiments. Furthermore, we expect the thesis work to enable a deeper understanding of the fundamental connections between symmetry, structure, and elementary excitations in condensed matter and material physics.