(376e) Machine Learning Guided Synthesis of Multinary Chevrel Phases for Tunable Energy Materials

The Chevrel phase (CP) is an inorganic family of molybdenum chalcogenides that have demonstrated great promise as advanced battery materials, artificial solid-electrolyte interphases and electrocatalysts for hydrogen evolution and CO₂ reduction. Multinary CPs comprise a vast compositional space with many thousands of theoretical members and varying these CP compositions can enable the precise tuning of electronic and thermodynamic properties that are relevant to energy applications. However, despite their promise as energy materials and the importance of understanding the range of synthesizable phases, CPs are underexplored with only ~100 synthesized compositions due to the challenge of identifying synthesizable phases. To catalyze the growth of this promising material class, we developed an interpretable machine-learned descriptor (H_δ) that rapidly and accurately estimates CP decomposition enthalpies (ΔH_d). H_δ was developed using ΔH_d values for 438 CP compositions computed with the accurate SCAN density functional. We developed and applied the new symbolic regression with intermediate feature trimming (SIFT) machine learning method, which provides an easy-to-use approach for rapidly developing accurate and interpretable chemical models. SIFT was used to generate over 560,000 descriptors of CP ΔH_d, of which H_δ yielded the highest accuracy. We applied H_δ alongside data-driven boundaries for ΔH_d that bracket stable and persistently metastable materials to identify over 2,000 CP compositions that are predicted to be synthesizable. 5/5 novel CP tellurides attempted from this set were experimentally synthesized, doubling the known space of this CP subset. We expect H_δ to catalyze the discovery of new CPs and to facilitate further development of CP design criteria that correlate composition and functionality in relevant energy applications. Furthermore, our computational approach for developing screening tools that enable the rapid identification of synthesizable materials is transferrable to other materials families to accelerate their discovery.