Discovery of Superionic Solid-State Electrolyte for Li-Ion Batteries via Machine Learning

Li-ion solid-state electrolytes (Li-SSEs)
hold promise to solve
critical issues related to conventional Li-ion batteries (LIBs), such
as the flammability of liquid electrolytes and dendrite growth. In
this study, we develop a platform involving a high-throughput screening
process and machine learning surrogate model for identifying superionic
Li-SSEs among 19,480 Li-containing materials. Li-SSE candidates are
selected based on the screening criteria, and their ionic conductivities
are predicted. For the training database, the ionic conductivities
and crystal systems of various inorganic SSEs, such as Na SuperIonic
CONductor (NASICON), argyrodite, and halide, are obtained from previous
literature. Subsequently, a chemical descriptor (CD), crystal system,
and number of atoms are used as machine-readable features. To reduce
the uncertainty in the surrogate model, the ensemble method, which
considers the two best-performing models, is employed; the mean prediction
accuracies are found to be 0.887 and 0.886, respectively. Furthermore,
first-principles calculations are conducted to confirm the ionic conductivities
of the strong candidates. Finally, three potential superionic Li-SSEs
that have not been previously investigated are proposed. We believe
that the platform constructed and explored in this work can accelerate
the search for Li-SSEs with satisfactory performance at a minimum
cost.