Real HF-trained H2 potential,three native MD interfaces: LAMMPS22Jul2025 update6/DeePMD3.2,ASE3.29,private GPUMD5.9.1 with optional USE_DEEPMD.
Use an ordinary account: bash analyze.sh;bash plot.sh;bash run.sh (P100) or bash run.sh v100. One Slurm GPU job builds a private DP-only GPUMD and executes six short NVE runs;no shared module changes,no new potential training. GCC13.3,CUDA12.9,OpenMPI5.0.10rc2;PyTorch2.12;two CPU threads for compilation and one MD thread.
Two timesteps0.05/0.1fs,same100fs duration,same initial stretch. Velocity units differ: LAMMPS Angstrom/ps,GPUMD Angstrom/fs,ASE converted internal units. GPUMD first frame0.5fs;do not fabricate a zero-time native frame.
Check totalenergy,coordinates,forces and timestep sensitivity before production. This small fixed HF basis model is not a transferable production H2 potential;no thermal sampling,NQE,transport or performance claims. GPUMD source license retained under gpumd-source;no executable binaries in ZIP.
