Predict Microstructure-Property Relationship of Steels and High Entropy Alloy Formation Using Machine Learning

Metallurgy and material design have thousands of years’ history and have played a critical role in the civilization process of humankind. Composition, processing, microstructure, and materials properties are the four cornerstones in materials research. The traditional trial-and-error method has been unprecedentedly challenged in the modern era when the number of components and phases in novel alloys keeps increasing, with high-entropy alloys as the representative. New opportunities emerge for alloy design in the artificial intelligence era. For the first part of the talk, we will present a successful machine-learning (ML) study using variational autoencoder and a regression model to identify the microstructure images with eye-challenging morphology for a number of martensitic and ferritic steels. The challenge in differentiating performance with respect to microstructure lies in the extreme similarity of these steel images, where differences seem insignificant to the viewer. The success of the approach is a result of highly optimized neural network structures and fine-tuned parameters therein. The model clearly identified the key role of several elements in the prototypical 9% Cr steel through the formation of the different features of the martensite steels. As such, a new inverse alloy design method is proposed based on neural networks. It demonstrates a systematic approach to design new alloys.<br/><br/>For the second part of the talk, we will present our research progress in predicting solid solution formation to accelerate high entropy alloys research. In the literature various empirical rules are proposed to predict the formation of single-phase solid solution, but many are based on very small datasets and hence are of very limited predictability. In this project, we perform a machine-learning study on a large dataset consisting of 1252 alloys, including binary and high-entropy alloys, and we achieve a success rate of 93% in predicting single-phase solid solution. The present ML results suggest that the molar volume and bulk modulus are the most important features, and accordingly, a new physics-based thermodynamic rule is constructed. The new rule is nonetheless slightly less accurate (73%) than the ML algorithm but employs only the elemental properties and is thus convenient in applications. Finally, the advantages and pitfalls in applying high-throughput screening and ML versus CALPHAD calculations will be discussed.