Exhaustive and informatics-aided search for fast Li-ion conductor with NASICON-type structure using material simulation and Bayesian optimization

Abstract

Currently, NASICON-type LiZr2(PO4)3 (LZP)-related materials are attracting attention as solid electrolytes. There are experimental reports that Li-ion conductivity can be improved by doping a small amount of Ca or Y into stoichiometric LZP. In previous studies, doping with only one element having a narrow search space has been attempted, and thus, further improvement of the Li-ion conductivity is conceivable by using multi-element doping. When multi-element doping is attempted, because the search space becomes enormous, it is necessary to evaluate the Li-ion conductivity using a low-cost method. Here, force-field molecular dynamics using a bond valence force field (BVFF) approach was performed to evaluate the Li-ion conductivity. We confirmed that the Li-ion conductivity of stoichiometric LZP derived from BVFF (6.2 × 10−6 S/cm) has good agreement with the first principle calculation result (5.0 × 10−6 S/cm). Our results suggest that the Li-ion conductivity can be further improved by simultaneously doping LZP with Ca and Y [6.1 × 10−5 S/cm, Li35/32Ca1/32Y1/32Zr31/16(PO4)3]. In addition, Bayesian optimization, which is an informatics approach, was performed using exhaustively computed conduction property datasets in order to validate efficient materials search. The averages for Bayesian optimization over 1000 trials show that the optimal composition can be found about seven times faster than by random search.