Machine learning guided discovery of trivalent high entropy oxides

Since their discovery in 2015, high entropy oxides (HEOs) have attracted attention for emergent properties enabling diverse technological applications. Initial attempts at the discovery of new HEOs followed the intuition from the alloys field, where combinatorial arguments suggested the existence of “literally billions” of new materials. The experimental reality in HEOs has proven more challenging. Their stability cannot be straightforwardly predicted based on chemical intuition alone due to the complex interactions between the many chemical constituents. Computational insights are crucial for narrowing the phase space, after which synthesis techniques can be tailored to identify the necessary conditions for stabilizing new HEOs.

 

In this work, we employ machine learning interatomic potentials (MLIPs) to study the synthesizability of a novel collection of trivalent cations. From nearly 500 possible compositions, we identify the 16 most promising candidates, which we then experimentally validate with solid state and combustion synthesis. Our experimental results were then organized into four categories that encompass the possible outcomes of an attempted HEO synthesis: (i) a mixture of competing phases, (ii) redox reaction, (iii) cation ordering, and (iv) the successful synthesis of an HEO. We conclude that higher entropy oxides are in a far more delicate energetic balance than initially believed and that machine learning approaches can effectively guide us to the "needle in the haystack".