Convolutional neural network based methodology for flexible phase prediction of high entropy alloys

Abstract

ABSTRACT The advent of high entropy alloys has created a design space that is unfeasible to explore solely through experimentation, thus necessitating the use of computational methods, such as artificial neural networks. This study proposes a new architecture utilizing a convolutional layer to extract relevant elemental features from the alloy composition without limiting the model to specific elements by treating the elemental composition of the high entropy alloy in a similar manner to a row of pixels, using relevant elemental properties in lieu of color values. This convolutional model was able to predict the crystal structure of solid solutions in a high entropy alloy composition with an accuracy of 89.3% in a six-way classification and the formation of intermetallics with an accuracy of 91.4% during holdout validation, while being capable of accurately predicting the primary crystal structure of high entropy alloys containing elements the model was not trained on. Significant space still exists for further experimentation and improvement with this methodology, including augmenting the available datasets and implementation of additional convolutional layers. Due to the limited interpretability of the model architecture, care should be taken when inferring trends from the model prior to validation through experimental, or well-established computational methods.