Date of Award
Fall 2025
Document Type
Open Access Dissertation
Department
Mechanical Engineering
First Advisor
Ming Hu
Abstract
Machine learning (ML) has gained tremendous attention as a potent and robust approach to discover and explore functional materials in a broad materials space. ML has been able to utilize atomic-level, chemical composition, and structural features of various materials and demonstrated massive success in predicting a wide variety of material properties with high accuracy comparable to density functional theory (DFT) calculations. Those properties predicted by ML include, but are not limited to optical, mechanical, thermal, and electronic properties. In the past decade, most properties tackled in previous ML studies were discrete, single-valued materials properties predicted. However, some materials properties that are remarkably and equally significant to study are continuous functions and have spectral-like formatting (i.e., not a single value) such as phonon density of states (DOS), the Eliashberg spectral function, and thermoelectric properties. Continuous and spectral properties pose several inherent challenges for ML, such as the sensitive requirements of the data smoothening and the existence of sharp peaks. Moreover, spectral properties usually have different maximum values and different cutoffs, making the length and/or resolution of the data non-uniform across various materials, and therefore it is hard to train a ML model for such spectral materials properties. Moreover, some other materials properties can be difficult to predict directly and must be normalized such as crystal orbital Hamilton population (COHP) and crystal orbital bond index (COBI). The normalized -ICOHP and normalized ICOBI can then be utilized as descriptors for extreme lattice thermal conductivity (LTC) materials. We developed graph neural networks (GNN) models for predicting phonon DOS, Eliashberg spectral function, and chemical bonding descriptors and deployed the GNN models to discover novel interfaces for high heat dissipation applications, conventional superconductors, and extreme LTC materials, respectively.
Rights
© 2025, Mohammed Saif Ali Al-Fahdi
Recommended Citation
Ali Al-Fahdi, M. S.(2025). Deep Learning Models for Predicting Complex Materials Properties for Novel Materials Discovery in a Broad Materials Space. (Doctoral dissertation). Retrieved from https://scholarcommons.sc.edu/etd/8703