Skip to content

Parameterizing Complex Reactive Force Fields Using Multiple Objective Evolutionary Strategies (MOES): Part 2: Transferability of ReaxFF Models to C-H-N-O Energetic Materials

Summary

Part 2 of the authors’ MOES series asks whether Pareto-optimal ReaxFF and ReaxFF-lg parameter sets discovered for RDX can transfer to other C–H–N–O energetic crystals without hand-tuned per-compound refits. Multiple Objective Evolutionary Strategies (MOES) searches high-dimensional parameter spaces under several QM-derived objectives simultaneously, returning families of fits rather than a single weighted optimum. The paper first optimizes RDX-centric training targets, then screens candidate parameter sets with fixed-protocol LAMMPS molecular dynamics on RDX and on five additional energetic solids (HMX, PETN, TATB, nitromethane, TATP) using identical integration, thermostat, barostat, and stress-control stages spelled out in the Methods section.

Methods

Force-field training (MOES on ReaxFF / ReaxFF-lg). Multiple Objective Evolutionary Strategies (MOES) search high-dimensional ReaxFF parameter spaces under several QM-derived objectives simultaneously, returning Pareto families of parameter sets rather than a single weighted optimum for RDX-centric training data enumerated in the article (papers/ReaxFF_others/Rice_MOES_JCTC_2015.pdf).

MD application (crystal screening in LAMMPS). All post-optimization screenings use LAMMPS with a uniform 0.25 fs timestep on three-dimensional periodic boundary conditions (PBC) crystal cells for each explosive listed below. The staged protocol tabulated in the Methods section alternates NVE and NVT blocks (2.5 ps, 10,000 steps each), NPT blocks (7.5 ps, 30,000 steps), and N\(_s\)T anisotropic-stress control segments (25 ps, 100,000 steps) at 300 K for RDX, HMX, PETN, TATB, nitromethane, and TATP crystals. Thermostat and barostat damping constants are 0.05 ps and 0.5 ps, respectively, as stated in the paper. Supercell dimensions and initial crystal setups are listed in the same tables.

Static QM / DFT reference data: QM objectives feeding MOES follow the benchmarks cited in the article (see tables for targets).

Replica / electric field / AIMD production: N/A — not part of the screening workflow described above.

Findings

MOES identifies Pareto-efficient parameter sets that meet or exceed the baseline ReaxFF-lg crystal-density and structural targets for RDX under the screening metrics tabulated in the article. Two MOES-derived models match or outperform ReaxFF-lg across the six energetic crystals in the transferability battery, suggesting that multi-objective search can surface reactive FF parameters that are not narrowly overfit to a single explosive. The discussion frames MOES as a practical global optimizer for ReaxFF-class bond-order forms where manual weight tuning is fragile, while still warning that coverage of training chemistries bounds confidence outside the C–H–N–O set tested here and that loading conditions beyond the fixed-protocol crystal tests require independent validation.

Limitations

MOES search is expensive; transferability tests cover only the six C–H–N–O crystals in the battery. Energetic materials outside the training envelope require refitting or independent validation.

Relevance to group

Illustrates evolutionary multi-objective fitting for ReaxFF-class models on energetic chemistry—adjacent to broader reactive FF practice in the corpus.

Citations and evidence anchors