Year: 2024
Type: dissertation
Abstract: The development of inorganic crystalline materials has been significantly transformed by the integration of data-driven and artificial intelligence approaches.Leveraging extensive datasets generated through both experimental and computational methods, machine learning models have emerged as powerful tools for material discovery.Recent advancements in AI, particularly the integration of advanced models like Generative Pretrained Transformers, have further revolutionized the field by autonomously proposing, executing, and refining experiments with human-level precision.Despite these advancements, several challenges persist, including issues related to data quality, prediction accuracy, model interpretability, and the seamless integration of computational and experimental workflows.Addressing these challenges, this thesis aims to contribute to this endeavor by three different AI-driven inorganic materials studies.The first study presents a data-driven approach to predict synthesizability using the existing high-throughput databases.Utilizing the Inorganic Crystal Structure Database and the Materials Project database, this approach represents crystal structures through a Fouriertransformed crystal properties representation and employs a deep learning model to predict synthesizability via a synthesizability score.The synthesizability model, designed as a synthesizability filter, demonstrated accurate classification capabilities, achieving an overall precision of 82.6% and a recall of 80.6% in predicting ternary inorganic materials.Training the model on compounds uploaded to the MP before 2015 and testing it on materials added after 2015 resulted in a high true positive rate accuracy of 88.60% on the post-2019 test set, alongside a precision of 9.81%, suggesting that many newly added materials remain unexplored and possess significant synthesis potential.Additionally, materials from the post-2019 dataset with the highest synthesizability scores were identified, recommending them for future studies.The synthesizability model, as a validation filter, is beneficial for future material screening and discovery.The second study underscores the potential of predictive modeling in materials science through three distinct research efforts.The first effort utilized machine learning (ML) and Abstract ii data-driven techniques to predict the synthesizability of ternary oxides, effectively guiding the experimental synthesis of ABO3 compounds.The second effort applied ML and Bayesian optimization to design multi-property Fe-Co-Ni alloys, leading to the identification of compositions with superior magnetic and mechanical properties, surpassing those of traditional alloys.The third effort employed ML to predict the thermoelectric performance of materials, accurately estimating figure-of-merit (ZT) values and revealing promising new candidates for further experimental investigation.Collectively, these studies demonstrate the efficiency and effectiveness of predictive modeling in accelerating the discovery and optimization of advanced materials.The final study introduces the WyCryst framework, a novel approach for generating inorganic crystal structures by integrating symmetry-aware representations, machine learning models, and an automated density functional theory workflow.The WyCryst framework was rigorously validated through ground-state and polymorph crystal structure prediction tasks, successfully reconstructing known materials such as CaTiO₃, CsPbI₃, and CuInS₂.Additionally, the framework generated Wyckoff genes, leading to the discovery of eight novel ternary compounds, all confirmed to be dynamically stable via phonon calculations.This study highlights WyCryst's potential to revolutionize AI-driven materials discovery by producing high-validity, symmetry-compliant crystal structures that are both physically meaningful and experimentally feasible.In conclusion, these studies offer a novel perspective to the transformative potential of data-driven and AI methodologies in materials development, demonstrating their effectiveness in predicting synthesizability, accelerating novel material discovery, and generative design of inorganic crystalline materials.These advancements emphasize AI's critical role in advancing material science and driving innovation across diverse applications.
Author: Zhu, Ruiming
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4410047891