Effective sampling of chemical space enhances materials discovery: the case of superionic Li-ion conductor Li₇Si₂S₇I

Discovering new materials requires navigating vast compositional spaces—choosing which elements to combine and finding experimentally realizable compounds. These decisions benefit from quantitative guidance to improve success rates and accelerate discovery.

Machine learning now helps us finding chemical patterns in experimental data, enabling decision-aiding computational tools. In this talk, I will describe how these ideas are implemented in four complementary frameworks: PhaseRank¹, PhaseSelect², PhaseBO³ and LEAFs⁴. PhaseRank prioritizes combinations of elements based on synthetic accessibility; PhaseSelect identifies elemental contributions to functional properties such as superconductivity and magnetism; PhaseBO applies Bayesian optimization to efficiently explore compositional spaces; and LEAFs introduces local-environment-aware atomic features to guide substitutional design.

Together, these techniques enabled the discovery of a novel superionic Li-ion conductor, Li₇Si₂S₇I ⁵ , and its derivative Li₇Si_2–xGe_xS₇I, which preserves the parent structure while exhibiting enhanced low-temperature ionic conductivity⁶.

This workflow offers a practical route for exploring chemical space efficiently and linking predictive modeling directly with synthesis. It also illustrates how close integration between computational modeling and experiment is essential for uncovering structure–function relationships⁷ and advancing functional materials.