Held-Out Test and Production-Model Provenance for the Pt–C–O NEP Used in Graphene/Pt(111) Intercalation Simulations
Description
This dataset provides held-out testing and production-model provenance analysis for the Pt–C–O fourth-generation neuroevolution potential (NEP4) used in the manuscript “Species-Dependent Interfacial Aggregation and Local Graphene Lifting on Pt(111)”. It supplements the principal computational dataset at https://doi.org/10.17632/h7m39h94hk.1 and the additional analysis dataset at https://doi.org/10.17632/cy22xdbxd8.1. The original database contains 4,226 configurations divided into 2,961 training configurations, 844 validation configurations, and 421 previously unused held-out test configurations. The 844-configuration subset previously described as a test set is treated here as the validation set. Production-model provenance was assessed by re-evaluating five archived frames from each public CO- and O-intercalation molecular-dynamics trajectory using two candidate Pt–C–O NEP models. The distributed production model (SHA-256: 4203fb759da3404878ecbde319f76f59014280d3cf6d88897312349fbffff28e) reproduces the archived trajectory energies substantially more closely than candidate model 2a. The fingerprint errors are 0.0200 versus 19.5630 meV/atom for CO and 0.00775 versus 1.24616 meV/atom for O, identifying model 2 as the model consistent with both archived production trajectories. On the 421-configuration held-out set containing 71,703 atoms, the frozen production model gives RMSEs of 0.98395 meV/atom for energy, 72.6538 meV/Å for force components, and 45.9592 meV/atom for virial components. Species-resolved force RMSEs are 48.7622 meV/Å for C, 93.1835 meV/Å for O, and 86.7958 meV/Å for Pt. The 844-configuration validation replay gives corresponding RMSEs of 0.94363 meV/atom, 74.6506 meV/Å, and 46.0115 meV/atom. The package contains the production NEP model and training input, held-out test structures and prediction outputs, validation-replay summaries, trajectory-fingerprint structures and results, reproduction scripts, the recorded software environment, and SHA-256 checksums. These validation and held-out results assess interpolation within the sampled Pt–C–O configuration domain. They do not establish unrestricted transferability to arbitrary compositions, defects, temperatures, reaction pathways, or configurations outside the sampled domain.
Files
Steps to reproduce
1. Verify the downloaded archive using its external SHA-256 file and verify the extracted contents using SHA256SUMS.txt. 2. The frozen production potential is 01_production_model/nep.txt (SHA-256: 4203fb759da3404878ecbde319f76f59014280d3cf6d88897312349fbffff28e). To repeat the held-out prediction, copy 02_heldout_test/heldout_test.xyz to train.xyz, copy 01_production_model/nep_training_input.in to nep.in, add prediction 1, and execute the NEP prediction program. No model fitting or parameter optimisation should be performed during this step. 3. Score the generated energy_train.out, force_train.out, and virial_train.out files using 05_reproduction_scripts/score_nep.py. The expected held-out RMSEs are 0.983953 meV/atom for energy, 72.653764 meV/Å for force components, and 45.959181 meV/atom for virial components. Species-resolved force RMSEs are 48.762181 meV/Å for C, 93.183508 meV/Å for O, and 86.795763 meV/Å for Pt. 4. Obtain the 844-configuration validation subset from the principal computational dataset at https://doi.org/10.17632/h7m39h94hk.1. This subset was supplied under the legacy filename test.xyz in the original NEP workflow; it is distinct from the 421-configuration held-out test set supplied in the present record. Re-evaluation with the frozen model 2 should reproduce RMSEs of approximately 0.943626 meV/atom for energy, 74.650618 meV/Å for force components, and 46.011477 meV/atom for virial components. 5. Production-model provenance is documented in 04_model_identity/. Five archived frames from each CO and O trajectory were evaluated with the production model and candidate model 2a. The resulting energy-fingerprint tables are provided as fingerprint_CO.csv and fingerprint_O.csv, together with the selected structures. 6. The validation calculations reported here were performed using the GPUMD/NEP 4.9.1 executable on an NVIDIA H100 node and Python 3.12.9. The recorded executable paths and software environment are provided in 06_environment/software_versions.txt. These tests assess model behaviour within the sampled Pt–C–O database and should not be interpreted as validation for arbitrary configurations outside that domain.
Institutions
- Institute for Basic ScienceDaejeon, Daejeon