Date of Award

Spring 2026

Document Type

Open Access Dissertation

Department

Computer Science and Engineering

First Advisor

Jianjun Hu

Abstract

The accelerated discovery of novel functional materials is critical for advancing trans- formative technologies in energy storage, electronics, and catalysis. Current discovery strategies, whether experimental trial-and-error or DFT-based computational screen- ing, are fundamentally constrained by the limited size and diversity of existing materials databases and the difculty of building predictive models that generalize beyond known compounds. Despite signifcant progress in both experimental and computational approaches, two central challenges persist in modern materials science: generating large numbers of realistic, stable, and synthesizable materials outside existing databases, and developing machine learning models capable of accurately inferring structures and predicting properties under realistic conditions. Addressing these challenges requires not only new generative and predictive methodologies, but also principled strategies for evaluating model generalization in scientifcally mean- ingful ways. This dissertation addresses both challenges through a series of deep learning and machine learning frameworks spanning materials generation, crystal structure prediction, structural inference, and property prediction. In the frst topic, a difusion language model framework is proposed for the generative design of novel inorganic materials. By coupling composition generation with template-based crys- tal structure prediction, graph neural network relaxation, and DFT validation, the framework successfully identifes stable compounds with negative formation energies and thermodynamic stability near the convex hull, demonstrating the potential of difusion-based generative models to explore vast chemical design spaces. In the second topic, generative modeling is extended to the discovery of two-dimensional (2D) materials, a domain where experimentally realized compounds remain scarce. Using a transformer-based generative model combined with large-scale data mining and DFT validation, diverse and stable 2D material candidates are generated, signifcantly expanding the structural landscape of this emerging materials class. In the third topic, CondADiT, a composition-conditioned latent difusion framework, is introduced for crystal structure prediction directly from chemical composition. The model integrates a conditional variational autoencoder with a difusion transformer and a per-token atom-type conditioning mechanism, enabling site-specifc structural guidance during generation. Operating in a learned Euclidean latent space allows CondADiT to avoid specialized geometric modeling techniques, while achieving state-of-the-art match rates on multiple established benchmarks. In the fourth topic, DeepXRD is presented as a deep learning framework for predicting X-ray difraction spectra directly from chemical composition, enabling scalable inference of structural fngerprints without requiring expensive simulations or experimental measurements. This capability provides a com- plementary and practical tool for structural characterization in data-driven materials discovery pipelines. In the ffth topic, domain adaptation techniques are systematically evaluated for materials property prediction under realistic distribution shifts. Unlike conventional models that assume identical training and testing distributions, multiple domain adaptation strategies are assessed across diverse application scenarios, demon- strating signifcant improvements in predictive performance for out-of-distribution materials and enabling more reliable generalization to novel compounds. Collectively, these fve contributions establish a comprehensive and cohesive data-driven frame- work that integrates generative modeling, structure learning, and domain-adaptive prediction to accelerate the discovery of stable, synthesizable, and functionally diverse materials.

Rights

© 2026, Rongzhi Dong

Download

Share

COinS