Calorine4.0 real-cluster examples
Load module calorine/4.0. CPU code is serial; run bash run-cpu.sh in allocated CPU resources.
GPUNEP: submit sbatch gpu.sbatch; all16 cases use the allocated card, gpu_identifier_index=None.
Batch C predictions: sbatch batch.sbatch. Compatible GPUMD5.9.1 C model.
Si stock model angular n_max10 works in CPUNEP/GPUMD MD but nep5.9.1 training/prediction parser requires angular n_max<=8; do not edit weights to bypass.
No new model is trained here. setup_training creates input directories only.
7/2 Si data files retain historical names train/test: correlated teaching groups, not original stock model training set or independent test.
DFT E/forces/virial are real; offsets demonstrate energy reference only, not improved accuracy.
Stress ASE=eV/A3 tensile-positive, prediction stress file=GPa pressure-positive; conversions/order are explicit.
Elastic constants in GPa use the relaxed equilibrium cell. NEP information entropy is not thermodynamic entropy.
Reference data and images are in reference/ and top-level PNG. Run scripts regenerate working data outside reference/.
