Date of Award

Winter 2024

Project Type

Dissertation

Program or Major

Chemistry

Degree Name

Doctor of Philosophy

First Advisor

Jiadong Zang

Second Advisor

John Tsavalas

Third Advisor

Mrityunjay Kothari

Abstract

This dissertation presents an integrated artificial intelligence framework aimed at accelerating magnetic materials discovery by addressing key challenges in materials science: improving experimental characterization techniques, efficiently extracting structured data from scientific literature, developing comprehensive materials databases, and creating reliable predictive models.

We first develop a UNet-enhanced vector field electron tomography (VFET) approach for three-dimensional magnetic structure reconstruction. This method addresses the missing wedge problem in experimental data collection while maintaining computational efficiency, providing reliable experimental data crucial for materials research.

We introduce GPTArticleExtractor, a novel workflow leveraging large language models to automatically extract key information from scientific literature. Applied to 22,120 articles from the Journal of Magnetism and Magnetic Materials, this system demonstrated the effectiveness of AI-driven data extraction by generating a database of 2,035 magnetic materials with detailed structural and magnetic properties.

Expanding upon GPTArticleExtractor's methodology, we develop the Northeast Materials Database (NEMAD), a comprehensive experiment-based magnetic materials database containing 26,706 entries. By enhancing and scaling up our extraction approach, NEMAD consolidates chemical compositions, magnetic phase transition temperatures, structural details, and magnetic properties from multiple data sources, including our experimental data and extracted literature data.

Utilizing NEMAD's extensive dataset, we develop machine learning models for magnetic property prediction. Our classification model achieves 90\% accuracy in categorizing materials as ferromagnetic, antiferromagnetic, or non-magnetic. The regression models exhibit high accuracy in predicting Curie and Néel temperatures, with R² values of 0.86 and 0.85, respectively. These models successfully identify 62 promising ferromagnetic candidates with predicted Curie temperatures above 500 K and 19 antiferromagnetic compounds with predicted Néel temperatures exceeding 100 K.

This integrated framework demonstrates the potential of AI-driven approaches in accelerating materials discovery by combining enhanced experimental techniques, automated knowledge extraction, comprehensive database development, and accurate property prediction. The methodologies and tools developed in this work lay a robust foundation for future advancements in materials science research.

Download

Share

COinS