DISSERTATION DEFENSE
 

Author :  Sadman Sadeed Omee
Advisors: Dr. Jianjun Hu
Date: Aug 10, 2026
Time: 04:00 pm
Location:  2265 (M. Bert Storey Engineering and Innovation Center)
Link:   https://sc-edu.zoom.us/j/4997546955

Abstract
The discovery of novel materials is essential for advances in energy storage, electronics, catalysis, superconductivity, and many other technologies. However, the vast chemical design space and the high cost of experimental and first-principles methods such as density functional theory (DFT) make large-scale materials discovery challenging. This dissertation develops artificial intelligence (AI) and machine learning (ML) methods to accelerate materials discovery through advances in property prediction, model generalization, crystal structure prediction (CSP), and generative modeling.
In the first topic, we introduce DeeperGATGNN, a scalable graph attention neural network that overcomes over-smoothing through architectural innovations, enabling deeper networks and achieving state-of-the-art performance for materials property prediction. In the second topic, we present the first comprehensive benchmark of out-of-distribution (OOD) generalization in materials property prediction, providing a systematic evaluation of leading models under challenging OOD settings and offering insights into their ability to generalize beyond the training distribution in realistic materials discovery scenarios. In the third topic, we present ParetoCSP, a hybrid AI-evolutionary framework that combines ML interatomic potentials with age-fitness Pareto optimization to improve the efficiency of CSP. In the fourth topic, we extend this framework with ParetoCSP2, which enhances polymorphism prediction through an adaptive space group diversity control mechanism, improved structural initialization, and iterative relaxation, leading to more reliable recovery of multiple stable crystal structures. In the fifth topic, we introduce Diffhedron, an E(3)-equivariant diffusion model that generates crystal coordination polyhedra from chemical compositions, providing a physically grounded intermediate representation for future CSP and inverse materials design problems.

Overall, this dissertation contributes new predictive models, benchmark studies, search algorithms, and generative frameworks that collectively advance AI-driven materials informatics. These developments move the field closer to the long-term vision of autonomous discovery systems capable of reliably linking composition, structure, and properties to accelerate the design of next-generation materials.