Data-driven Models and Computational Frameworks for Physically Informed Design of Compositionally Complex Alloys

Abstract

Compositionally complex alloys (CCAs), containing large number of elements in high concentrations, represent an astronomical design space that can span more than 10100 possible compositions. The properties of CCAs can be improved significantly via compositional and microstructural tuning; but the traditional methodologies i.e., experimental alloy design and ab initio calculations are not suitable for efficient exploration of CCAs owing to their vast design space. Thus, machine learning (ML) has taken a center stage in recent years and various ML models have been reported for the exploration of CCAs. But these are often treated as a black box that offers no physical insights into the decision-making process of the trained models. In this thesis, we have developed interpretation frameworks, machine learning models and computational tools that enable targeted and physically informed exploration of CCAs. To address the black-box nature of ML models, two approaches have been implemented. Firstly, a neural network based ML has been reduced into a simpler and fundamentally interpretable mathematical model that can predict the probability of occurrence of FCC and BCC phases in CCAs. Secondly, a novel model-agnostic Compositional-Stimulus and Model Response (CoSMoR) framework has been developed to extract exact contribution of individual features. CoSMoR establishes the physical consistency of the nature of fit and provides material-specific specific insights into the decision-making process. We have also developed ML models for the prediction of short-range order (SRO) and hardness in CCAs and have validated them over a variety of complex alloy systems through comparison with ab-initio and experimental results wherein they reliably capture the linear, non-linear and non-monotonic changes in hardness and SRO as a function of alloy composition. We have also carried out experimental studies on CoCrNi ternary, CoCrCuNi quaternary and CoCrCuNi-M quinary alloys to probe the effect of strong ordering and clustering binary pairs on the overall phase evolution. The strong and contrasting binary pair interactions encountered in these alloys provide a good test bed for not only validating but for also finding the limits of the ML models. To support our experimental studies and to address the challenges faced in identification and quantification of phases during microstructural characterization of CCAs using SEM-EDS data (especially when phase contrast is missing in SEM images), we have developed EDS-PhaSe software that performs phase segmentation and analysis through quantitative analysis of EDS elemental maps. All feature generation programs, trained ML models and interpretation routines developed as part of this thesis have been packaged as an open-source Python library (MAPAL).