Stable GPU examples: existing VASP6.6.1-VTST NVHPC25.3 builds for P100 / V10016G, one MPI rank per GPU.
Optional ML example: VASP6.6.1 std + VTST revision214 vtstcode6.6.0 CPU build; Intel oneAPI2024.2 mpiifx, IntelMPI2021.13, sequential MKL and serialHDF51.14.6.
The dedicated mlfix variant resets the per-force-call ML cleanup flag and restores the actual incoming DFT energy after a surrogate loop. It is a distinct module, not the stock GPU VTST module.
Actual small-model ML examples exercised1 and2MPI; an independent normal public VASP GPU calculation confirmed the final geometry and force. OrdinaryforceCG also ran with the correctedCPUbinary.
The NN is an online local surrogate from this2Si example, not a transferable trained silicon potential. Use actual DFT convergence criteria for final structural results.
VASP source code, binaries, licensed POTCAR and full private outputs are not included in this website archive.
