## Date of Award

Summer 2023

## Document Type

Open Access Dissertation

## Department

Mechanical Engineering

## First Advisor

Ming Hu

## Abstract

Currently, heat transfer in many industries is the limiting factor for innovation, especially in the energy sector. For example, maximizing thermal conductivity of ceramic coatings in power plant devices improves the overall electrical to thermal energy ratio, whereas minimizing thermal conductivity is required for desirable heat-to-electricity conversion in thermoelectric devices. As such, rapid discovery of new materials with extreme thermal conductivity values is quintessential for the near-future deployment of current and developing energy applications.

The vibrational properties of crystalline materials are essential for their ability to conduct heat. Fundamentally, the restorative atomic forces of displaced atoms are sufficient to represent phonons, or quasi-particles of vibrations. As such, accurate ab initio calculators such as density functional theory (DFT) are standard for computing atomic forces and subsequent interatomic force constants (IFCs) for phonon properties, such as thermodynamic stability and lattice thermal conductivity (LTC). However, depending on the crystal symmetry and the order of the interatomic force constants, several tens to hundreds of simulations containing displaced atoms are required for a single crystalline material. In the context of high-throughput prediction, the costly and time-consuming nature of DFT renders this infeasible when facing tens of thousands of new materials for prediction.

This dissertation focuses on the development, application, and demonstration of machine learning-based models for rapid virtual evaluation and screening of materials for thermal applications. As an atomistic or “bottoms-up” approach, the Spatial Density Neural Network Force Field (SDNNFF) model proposed here predicts atomic forces which may be used in standard phonon calculators and Boltzmann Transport Equation (BTE) codes at significantly faster speeds than DFT. Specifically, atomic forces from reliable quantum mechanical simulations are trained in neural networks, demonstrating a < 10^{-2} eV/Å force accuracy which is on the similar level as pseudopotentials. Additionally, the speed of the SDNNFF relies on the atomic or input descriptors, providing at minimum a 10^{3} speed-up for systems containing > 100 atoms in comparison to first-principles DFT. Correspondingly, high-throughput thermal conductivity calculations are made possible for databases containing > 10^{5} theoretical materials.

To demonstrate database evaluation of LTC, the SDNNFF inputs are modified to accommodate an arbitrary number of elements, which is trained with a dataset spanning up to 63 elements and 16 structural prototypes in a single model. The abundant dataset size of 3×10^{7} atomic environments for training is made possible by the training on atomic forces, yielding more information per costly DFT run. Additionally, active learning techniques provide iterative improvement of the model with little to no human intervention. In total, 88,597 materials are evaluated for their thermodynamic stability, whereby the 34,432 remaining materials are queried for lattice thermal conductivity values. Other exotic properties, such as Weyl points, p-d orbital hybridization, and phonon hydrodynamics, are made possible from the predicted atomic forces. The properties of materials found here are also verified with existing experimental and computational works with agreeable accuracy.

## Recommended Citation

Rodriguez, A. D.(2023). *Development of Atomistic Machine Learning Approaches for Thermal Properties of Multi-Component Solids and Liquids.* (Doctoral dissertation). Retrieved from https://scholarcommons.sc.edu/etd/7498