Temperature-induced phase transitions and ionic conductivities of Li 2 B 12 H 12 and LiCB 11 H 12 were simulated with the use of machine learning interatomic potentials based on van der Waals-corrected density functional theory (rev-vdW-DF2 functional). The simulated temperature of order−disorder phase transition, lattice parameters, diffusion, ionic conductivity, and activation energies are in good agreement with experimental data. Our simulations of Li 2 B 12 H 12 uncover the importance of the reorientational motion of the [B 12 H 12 ] 2− anion. In the ordered α-phase (T < 625 K), these anions have well-defined orientations, while in the disordered β-phase (T > 625 K), their orientations are random. In vacancy-rich systems, its complete rotation was observed, while in the ideal crystal, the anions display limited vabrational motion, indicating the static nature of the phase transition without dynamic disordering. The use of machine learning interatomic potentials has allowed us to study large systems (>2000 atoms) in long (nanosecond-scale) molecular dynamics runs with ab initio quality.