Is There a Simple Descriptor to Predict Laves Phases?

Abstract

Laves phases AB₂, which represent the largest group of intermetallic compounds, have many applications as structural and functional materials, whose properties can be optimized through the tuning of solid solutions such as (A1,A2)B₂ or A(B1,B2)₂. Although they are known to be governed by size and electronic factors, there is no universal set of rules that is able to predict which arbitrary combination of elements will lead to Laves structures. Models have been recently developed that can predict Laves structures accurately based on conventional machine learning algorithms, but more interpretable models would be desirable. Through application of the sure independence screening and sparsifying operator (SISSO) method, modified using decision trees as the scoring function, simple descriptors based on elemental properties were sought to classify Laves vs non-Laves structures within a data set consisting of 534 binary and 3833 ternary experimentally known intermetallic phases reported in Pearson’s Crystal Data. A model based on a one-dimensional descriptor was proposed that depends on elemental properties of the A and B components, with the electron density at the boundary of the Wigner–Seitz cell for the B component playing an important role. This model gave an accuracy of 90% in predicting Laves vs non-Laves structures among binary and ternary phases. As a test of the model, the solid solubility limits for Dy(Ag_xAl_1–x)₂ and Er(Ag_xAl_1–x)₂ Laves phases were predicted and then experimentally validated through arc-melting reactions and structural characterization.