Computation and Data-Driven Methods Toward Superhard Materials Design

High hardness materials as measured through Vickers microindentation testing are indis- pensable for a myriad of industrial applications. The development of new high hardness ma- terials has traditionally relied on trial-and-error methods or empirically derived designing rules. This has significantly hindered the development of novel superhard materials. The complex relationship between crystal structure, composition, bonding, and hardness also remains difficult to elucidate and requires more insight. This contribution employs compu- tational methods to understand the fundamental mechanical properties of hard materials and machine learning methods to design predictive models that can capture the complex composition-structure-property relationship of high hardness materials. Moreover, an un- supervised learning model is designed to capture the underlying crystal-chemical patterns in existing data and provide new chemical knowledge from the learning process. First, the influence of chemical bonding is studied on the mechanical properties of earth-abundant transition metal borides through density functional theory. These results showed that bonding optimization through elemental substitution can effectively tailor a material’s mechanical properties. Then, a machine learning model is developed to map the complex relationship between chemical composition, applied load, and Vickers hard- ness (HV ). Large-scale screening is further enabled by this new method for high hardness materials by directly predicting HV at various applied loads. More than ten thermo- dynamically favorable compositions are identified to be superhard, proving the machine learning model’s ability to find previously unknown materials with outstanding hardness. Then, the influence of temperature on hardness is studied by adding the measurement temperature as an additional variable in a supervised learning model. The reported model showed excellent performance in estimating the hardness decrease at elevated tempera- ture, which is extremely useful for identifying thermally-robust hard materials. Finally, this work overcomes the main limitation of supervised learning–sparse and small training datasets. This is one of the main reasons machine learning struggles to find materials that outperform the state-of-the-art. Unsupervised learning in the form of an anomaly detection framework is developed using an autoencoder architecture. The model is trained on known crystal structure data learning to distinguish normal crystal-chemical patterns from anomalies. Further analyzing the structural factors that contribute to high anomaly scores revealed that due to the unexpectedly short bond lengths in superhard materials, they are statistically considered anomalies among all inorganic materials. Together this dissertation provided insights on chemical bonding of structural materials, demonstrated highly efficient and accurate predictive machine learning models that captured the complex relationship between external factors and hardness, and provided advanced unsupervised anomaly detection methodologies to further understand rare materials’ properties.