ReaxFF Parameter Set for Boron Clusters and Icosahedral Boron Crystals: Comparison with Density Functional Theory and Machine Learning Potentials
Corpus note
The ingested PDF is an ACS galley/manuscript proof (Comer_Liu_Boron_JPC_2025_galley.pdf), not necessarily the final version-of-record layout.
Summary¶
The work refits a boron-focused ReaxFF parameter set so cluster relative energies align better with DFT, using a training set centered on B\(_{80}\) motifs (including core–shell and “Pouch” structures from earlier simulations). The refined set improves agreement on a broader test set of boron clusters (8–103 atoms), compares favorably to several machine-learning interatomic potentials on the same B\(_{80}\) ranking task, and is exercised in larger-scale LAMMPS simulations of crystallization from supercooled liquid boron and of boron solubility at a boron–molten nickel interface.
Methods¶
- DFT reference data: Periodic cluster calculations used VASP with PBE (and related setups described in the article); plane-wave cutoff 520 eV, Γ-only k-point sampling, Fermi–Dirac smearing (0.2 eV), BFGS optimization with force threshold 0.02 eV/Å. Cluster coordinates were prepared with ASE in 20×20×20 Å\(^3\) cells with 10 Å vacuum padding.
- ReaxFF training and versions: An initial “Version 1” set was refit to penalize spurious low-energy boron cluster motifs, leading to Version 2 and a final Version 3 (B-v3) trained with explicit handling of the “Pouch” geometry and its DFT energetics. Relative energies reported for B\(_{80}\) structures use geometry optimization under each model before comparing cluster energies.
- Reactive MD (LAMMPS): Simulations used LAMMPS (23 Jun 2022) with ReaxFF and GPU-accelerated routines. Time step 0.25 fs. Isolated clusters: NVT with Langevin thermostat (200 fs damping). Melts and continuous solid systems: NPT with Nosé–Hoover thermostat and barostat (100 fs time constant). Trajectories saved as DCD every 1000 steps; visualization and analysis used VMD and in-house Tcl scripts (deposited as noted in the paper).
- Seed-crystal growth tests: β-rhombohedral boron supercells with restrained seeds were melted (3600 K), then cooled to 1300–1700 K at 1 atm for 10–30 ns; additional runs explored ~10 and ~24 GPa and larger supercells as described in the article.
- Boron–Ni solubility estimate: Canonical (NVT) simulations placed solid boron against a molten Ni(111)-derived slab; boron fraction was estimated in a slab-centered region (\(|z-z_{\mathrm{Ni,COM}}|<5\) Å) versus time.
Findings¶
- The final B-v3 parameter set reproduces the correct relative ranking of the key B\(_{80}\) training structures compared to DFT, including a large penalty for the spurious “Pouch” motif that earlier parameter sets favored.
- On the training trio of B\(_{80}\) structures, B-v3 and PFP v7.0.0 were among the few models that matched DFT ranking; B-v3 achieved the lowest RMS error versus DFT among the empirical and MLIPs surveyed in the comparison table.
- Extended melts and growth simulations show more icosahedral character with the refined parameters than with the prior boron ReaxFF parameterization.
- Boron solubility in contact with molten nickel from the refined parameters is closer to experimental expectations than the older boron parameter set, which underestimated solubility strongly.
Limitations¶
Boron parameterization is cluster- and B\(_{80}\)-centric; transfer to all bulk allotropes and impurity chemistries is not guaranteed. Proof/galley PDF pagination may differ from the journal version. GPU LAMMPS timings and thermostat choices affect observable nucleation pathways in long melts.
Relevance to group¶
Jeffrey Comer, Adri van Duin, Bin Liu, and collaborators refine boron ReaxFF for icosahedral boron crystallization and metal contact scenarios—adjacent to materials synthesis modeling in the corpus.