Skip to content

Neural network and ReaxFF comparison for Au properties

Summary

Machine-learned interatomic potentials and physics-based reactive force fields are both trained against Kohn–Sham DFT data, but they differ in flexibility, data efficiency, and cost per MD step. This International Journal of Quantum Chemistry paper uses gold as a deliberately simple, screened single-component metal to compare ReaxFF against Behler–Parrinello neural-network (BPNN) potentials across bulk, surface, cluster, and pathway (e.g., NEB) regimes. The study asks how much training data each model class needs to be reliable, where each breaks down, and what trade-offs appear when the goal is broad coverage of structural chemistry rather than a single narrow application. The overarching motivation is practical: practitioners need guidance on when a compact reactive FF is sufficient versus when a high-capacity regressor is warranted, and what validation splits are meaningful across regimes.

Methods

MD application (atomistic dynamics)

This paper is a potential-energy-surface benchmarking study: headline data are static KS-DFT energies/forces on bulk, surface, cluster, and NEB configurations used to train and test ReaxFF and Behler–Parrinello neural networks, not reported finite-temperature molecular dynamics production trajectories.

  • Engine: N/A — no LAMMPS/GROMACS/CP2K molecular dynamics trajectories are reported as primary results; VASP supplies DFT reference forces/energies.
  • System size & boundaries: training cells vary by task (bulk EOS, surfaces, ≤126-atom clusters, NEB images); periodic boundary conditions (PBC) follow the VASP setups described in Methods, but there is not one fixed MD supercell for the whole paper.
  • Ensemble / timestep / thermostat / barostat / duration (production MD): N/A — not in scope because the contribution is DFT-labeled fitting rather than NVT/NPT/NVE MD sampling.
  • Electric field: N/A — not used. Replica / enhanced sampling: N/A — training uses relaxation pathways and climbing-image NEB images, not umbrella sampling or replica-exchange molecular dynamics.

Force-field training (ReaxFF)

  • Parent FF / elements: ReaxFF for Au (reactive bond-order formulation as described in the article’s background).
  • QM reference data: Kohn–Sham DFT computed in VASP using PBE-GGA and PAW pseudopotentials; Monkhorst–Pack k grids at least 14×14×14 for a single Au atom in the primitive ground-state reference; plane-wave cutoff ≥ 300 eV targeting ≤5 meV/atom energy convergence; relaxations to <0.05 eV Å⁻¹ (Methods section).
  • Training set scope: 9,972 KS-DFT calculations total (905 bulk, 1,022 surface, 8,045 cluster), mostly intermediate relaxation images; 896 additional local minima / NEB images.
  • Optimization: Monte Carlo Force Field optimization (MCFFopt) in ADF for ReaxFF parameter searches.
  • Best-fit data count (reported): A strong ReaxFF model trained from 848 points performs well on bulk/surface held-out data but remains comparatively weak for ≤126-atom clusters in their tests; expanding training can overfit and degrade transfer.

Force-field training (Behler–Parrinello neural network)

  • Software: AMP (Peterson/Khorshidi) integrated with ASE for BPNN training; trained-parameter details stored as JSON in Supporting Information (as stated in the article).
  • Training set scope: Example large fit uses 9,734 KS-DFT calculations from the same pool.
  • Performance claim: BPNN matches or exceeds ReaxFF across the tested bulk/surface/cluster regimes in their benchmarks, with higher per-step computational cost in the implementation discussed.

Static QM / DFT (KS reference generator)

Same VASP/PBE/PAW settings as above; climbing-image NEB used where transition states are reported (unless otherwise noted in the Methods text).

Findings

  • ReaxFF vs BPNN trade-off: ReaxFF can reach strong bulk/surface accuracy with far fewer training points, but is much weaker on small Au clusters (≤126 atoms) in the reported errors.
  • Overfitting warning: For ReaxFF, adding more training data does not monotonically improve held-out performance and can hurt transfer—an explicit caution about automated reactive fits on heterogeneous datasets.
  • BPNN scaling: BPNN benefits from large training coverage and tracks or beats ReaxFF across regimes in their tests, at higher MD-step cost in the studied implementation.
  • System choice rationale: Au is used partly because screened metallic interactions reduce the role of long-range electrostatic extremes, making the comparison a cleaner probe of short-ranged PES modeling between physical and machine-learned potentials (with the article’s caveat that polar/ionic systems add further requirements).

Limitations

Transfer to oxides, electrolytes, or charged interfaces requires different datasets and validation. Reported timings and cost models depend on NN architecture and software path.

Relevance to group

Benchmark culture for ReaxFF versus neural potentials on metals—adjacent to MLIP discussions in reactive MD communities.

Citations and evidence anchors