Date of Award

Fall 2024

Document Type

Open Access Dissertation

Department

Mechanical Engineering

First Advisor

Ming Hu

Abstract

The quest for materials with extraordinary properties has been a longstanding endeavor in material science and engineering, driving future technological advancement. However, the discovery of such materials is non-trivial. Recent advancements in computational methods, particularly the integration of machine learning (ML) techniques with density functional theory (DFT), have opened new avenues for accelerating the discovery of materials with exceptional and extreme properties. This dissertation focuses on developing a synergistic approach and workflow combining ML and DFT to identify materials with properties that are pushed beyond current limits, using lattice thermal conductivity (LTC) as a case study of the workflow.

We begin by applying various machine learning (ML) algorithms, including Random Forest, Categorical Boosting (CatBoost), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), the Atomistic Line Graph Neural Network (ALIGNN), the Orbital Graph Convolution Neural Network (OGCNN), and the Global Attention Graph Neural Network (deeperGATGNN), to predict and classify materials with low or high lattice thermal conductivity (LTC). This multi-algorithm approach balances the high accuracy of quantum-level calculations with the computational efficiency of classical methods. The iterative loop of training, evaluation, recommendation, and verification continues until no additional materials are identified for recommendation from the screening pool.

A critical step in this workflow is the significant expansion of the unknown or untested materials pool. Traditional crystal structures for screening are sourced from existing databases such as the Materials Project and the Open Quantum Materials Database, limiting the screening pool. This dissertation addresses this limitation by generating new dynamically stable crystal structures using the graph theory-assisted universal structure searcher (MAGUS), combined with physical and chemical features learned from existing materials. This approach aims to discover more stable structures with extremely low LTC.

Additionally, we demonstrate the effectiveness of this methodology in other property domains, such as mechanical hardness and heat capacity. For example, using a deeperGATGNN model, we predicted the heat capacity of 32,026 structures, identifying 22 materials with high heat capacity, including MnIn2Se4, which exhibits an exceptionally high heat capacity exceeding that of Dulong-Petit limit at room temperature. These comprehensive efforts advance our understanding of material properties and hold significant potential for practical applications in energy conversion, storage, and advanced manufacturing.

Rights

© 2025, Joshua Ojih

Download

Share

COinS