A machine-learned interatomic potential for defects investigation in lead-free double perovskite Cs2NaYbCl6

We derive and validate a machine-learned interatomic potential, based on the Chebyshev Interaction Model for Efficient Simulation, for the lead-free double perovskite Cs2NaYbCl6, with special emphasis on native defect behavior. Starting from Density Functional Theory (DFT) reference data, we construct and benchmark several ChIMES parametrizations, varying body-order expansions and cutoffs, against DFT for equilibrium lattice constants, Birch-Murnaghan bulk moduli, and lattice thermal conductivity. The optimal parametrization reproduces with high accuracy the lattice constant, bulk modulus, thermal conductivity, and chlorine-vacancy formation energy while limiting computational workload, with an efficient scaling up to 103/104 atoms. We finally benchmark the potential by comparing radial distribution functions from molecular dynamics at 300 K. Long NVE runs on pristine and Cl-vacant supercells (up to 500 ps) confirm excellent energy conservation.