Machine Learning for Novel Ternary Materials Discovery

Abstract

First principles codes based on DFT are now sufficiently accurate and efficient that they can be used in the design of novel materials with specifically selected properties. Doing so requires checking whether a compound considered is chemically stable or not, which in turn involves constructing an appropriate convex hull. While such an approach is valid for identifying new, synthesizable materials, it demands numerous DFT calculations to ensure reliability. This work aims at accelerating this task through the use of machine learning interatomic potentials (MLIAPs) as screening agents. The process is facilitated by using pre-existing data on large material repositories to form the training sets of the models. Owing to the relative wealth in the number of binary systems compared to the ternaries, the former provide an ideal and extensive database for this training. In contrast, the space of ternaries, being only sparsely examined, forms fertile ground for exploration. In a first part of this work, an ensemble of spectral neighbour analysis potentials (SNAPs) is trained on binary data of the Ag-Au-Cu system taken from the AFLOWlib repository. The model is tested on different datasets composed entirely of ternary intermetallics. It is shown that an accuracy below 30 meV/atom can be achieved for alloys in their equilibrium structures, sufficient for an effective screening model. The MLIAPs are however unable to perform relaxation due to their poor energy predictions on out-of-equilibrium structures. Since suitable error metrics, capable of pinpointing unrelaxed structures, are verified, the devised model can be used in a high-throughput screening setting, in which candidates are physically sound compounds. In the follow-up study, this surrogate to DFT is incorporated into a workflow aimed at constructing reliable, DFT-level ternary convex hulls. This is achieved by two means. Firstly, the prototypes used as candidate ternary compounds are built from the structures of the low-enthalpy alloys of the binary subsystems. These form reasonable guesses for equilibrium structures, owing to the close similarity between binary and ternary alloys for transition metals. Secondly, measures are taken in order to increase the robustness of the screening process. These notably involve undertaking partial ionic relaxation, driven by SNAP, as well as an assessment of the reliability of the predictions made, through the use of an error metric. The final workflow developed is tested on Ag-Au-Cu and Mo-Ta-W, and is capable of identifying novel ternary compounds, absent from AFLOWlib, and thus produces DFT-accurate ternary convex hulls. This is achieved by probing a large number of candidates and focusing all the heavy ab initio calculations on the most promising candidates. In a final section, the recently introduced M3GNet universal force-field is inserted into the workflow. This enables higher accuracy and throughput, as exemplified by the better convex hulls obtained and the larger number of compounds tested. It is shown how this M3GNet workflow can be used to identify promising regions of ternary convex hulls, even for magnetic systems.