Insights into Structure-Property Relationships via First Principles and Machine Learning Approaches

Abstract

The increasing concerns about a sustainable future demand accelerated design and development of new materials with enhanced properties. The process of material development immensely depends on the structure-to-property (SP) relationship. However, obtaining a property from a material or finding a material with a desired property is not straightforward. It requires a thorough understanding of the chemical and physical properties of the material. Exploring the available large and diverse chemical space in alloys following the traditional route would take decades. As an alternative to the exhaustive traditional approaches for alloy developement, data science has emerged as a powerful tool to comprehend the structure-property relationships in alloys. This thesis focuses on different design strategies for optimizing structural materials and metal contacts using first principles-based density functional theory (DFT) and machine learning (ML) approaches. We attempt to address the challenges in structural materials for high mechanical strength and failure lifetimes through different material representations. We developed an extensive experimental database comprising compositions, heating conditions, and processing parameters through literature survey. Using these descriptors, we proposed an interpretable optimization framework for enhancing the yield strength and ultimate tensile strength in titanium alloys. We presented insights on developing new alloys with enhanced mechanical performance through the constrained optimization framework. Further, we addressed the ML predictions of failure lifetime in titanium alloys using the compositions and processing conditions. We efficiently identified the parameters guiding the creep and fatigue failures through explainable AI methods. Using ML, we could resemble the empirical relationship between the stress-number of cycles (S-N) curve in fatigue and the general equation of creep life. Despite considering the features mentioned above for ML, we next aim to unveil the role of structural parameters in predicting properties. We established the SP relationship by mapping microstructures to the target properties in nickel and cobalt-based superalloys. Relying on experimental data from the literature, we developed a database to predict the Vickers' hardness using image processing and supervised learning. We implemented different image processing tools to identify the distribution and geometry of phases present in the superalloy microstructures. We developed a highly accurate ML model for predicting Vickers hardness by integrating the microstructural features with the compositional and processing conditions. To further elevate the unique structural representations at the atomistic scale, we aimed to tackle the configurational and compositional diversities in high entropy alloys (HEAs). We developed a material-agnostic, explainable, high-fidelity descriptor based on crystal graph representation, which integrates the chemistry and atomic environment to predict the formation likelihood of stable phases. The designed descriptors could identify and elucidate the key factors governing phase formation. We also studied the thermodynamics and kinetics of metal/metal contact for copper interconnects. We reported a stable liner metal for enhancing the electromigration reliability in copper interconnects in integrated circuits. The thermodynamic stability and diffusion kinetics of copper/liner contacts is examined using DFT and experiments. Our research findings pave the way to overcoming some critical challenges related to structural materials and copper interconnects. Further, our approach could accelerate the design and discovery of new materials at a faster and cheaper scale.