Rapid screening of high-entropy alloys using neural networks and constituent elements

Particularly over the past year, there has been a significant surge in machine learning (ML) methods to predict new High-Entropy Alloys (HEAs). Despite considerable achievements from these attempts, their models require datasets built on thermodynamic and Hume-Rothery (HR) features of the alloys. Often, some of these features are taxing to obtain and involve substantial estimations. Considering that an alloy’s composition is always known for certain, we present the idea of using compositional (atomic percentage) data with a simple neural network to predict HEA phases. This is to encourage the use of ML and enable the average researcher to rapidly screen potential HEA compositions. We also explore a neural network which uses HR features along with the atomic percentage data and compare the two models. The neural network working with compositional data only (NN1) achieved an average testing accuracy of 92%, whereas the neural network working with HR features as well as compositional data (NN2) achieved an average testing accuracy of 90%. That NN1 outperforms the vast majority of models of the present date has significant implications, as it shows that the complete abandonment of predictive features is not only possible but also advantageous. To validate the models, we use NN1 and NN2 to predict the solid-solution window of the AlxCrCuFeNi system. Both networks predicts that beyond x = 1.4, the system is predominantly intermetallic. This was confirmed by arc-melting the x = 1.0, x = 1.3, x = 1.5 and x = 2 samples and studying their microstructures using SEM imaging.