Feature-assisted Predictability of Machine Learning to Design Materials with Unique Mechanical Properties

The mechanical properties of materials, spanning from nanoscale to continuum scale, are critical for various engineering applications. At the nanoscale, these properties are heavily influenced by microstructure and defects, which affect overall performance. While single-crystal models provide insight into fundamental material behaviors, designing materials with exceptional elastic properties at the atomic scale remains challenging. One effective approach to overcoming this challenge involves gaining atomic-level insights into elastic properties through atomistic modeling and data-driven methods. Atomistic modeling focuses on understanding atomic interactions, but it is computationally expensive and impractical for large-scale applications. In contrast, data-driven methodologies, particularly AI and machine learning (ML), offer a promising alternative, enabling insights into material behaviors from experimental or simulated data. Despite the growing use of ML models in material discovery, challenges remain, including limited and heterogeneous data, optimal representation of chemical features, and the need for standardized predictability measures. This work addresses these challenges through rigorous validation protocols across various ML model architectures, aiming to predict elastic properties effectively. Within a robust feature space, the study highlights the statistical limitations of predictability and reveals how different features influence elastic properties. Findings indicate that while physical attributes are essential, the local chemical and structural environment is equally crucial in determining material behavior. Rather than relying solely on predetermined features, this work introduces a feature-directed learning approach using advanced architectures like Graph Neural Networks (GNNs). This approach not only enhances predictive performance but also facilitates the transfer of domain knowledge, improving GNN models' learning capabilities. The refined model enables exploration of new candidates for materials with exceptional mechanical properties, such as super hardness. Overall, this study contributes to materials science by integrating advanced ML techniques with traditional material characterization, paving the way for developing novel materials with tailored mechanical properties for future engineering applications.