Skip to content

Efficient global optimization of reactive force-field parameters (ogolem / serial ReaxFF preprint)

Corpus note

The local PDF is a preprint / in-press manuscript (J. Comput. Chem., 2015 submission metadata on the cover page). The published article is curated as [[2015dittner-venue-efficient-global]] (DOI 10.1002/jcc.23966); prefer that entry for authoritative pagination and final title.

Summary

This ingested PDF corresponds to the same research line as the published Journal of Computational Chemistry paper on efficient global optimization of reactive force-field parameters: it presents ogolem, an evolutionary or genetic-algorithm framework with thread-level and MPI-level parallelism, extended to complex ReaxFF training sets that couple many structures and observables. The manuscript argues that global meta-heuristic search can reach high-quality ReaxFF parameter sets with less manual intervention and favorable parallel scaling compared with serial multistart local optimization workflows common in older practice. Conceptually, the code treats each fitness evaluation as an embarrassingly parallel task across candidate parameter vectors, which maps well to clusters running LAMMPS or PuReMD energy calls. The preprint therefore documents an engineering effort to industrialize reactive force-field fitting beyond ad hoc scripts.

Methods

Force-field training (global optimization workflow)

Parent potential: ReaxFF parameter vectors for reactive systems. QM / reference data: DFT (and related QM) energies/forces on training structures supply the objective, alongside experimental benchmarks where included in the manuscript’s training set. Optimization: ogolem implements genetic-algorithm / evolutionary global parameter optimization (selection, crossover, mutation) with least-squares-style fitness reductions versus the reference data. Training set: multi-structure ReaxFF fitting sets coupling many geometries and observables (Sec. 2.2 narrative in the preprint).

MD application (benchmark energy evaluations)

The manuscript positions ogolem as driving repeated energy and force evaluations on ReaxFF systems—typically via molecular dynamics or energy minimization on periodic supercells whose atom counts and stoichiometry follow each benchmark’s training geometry (full tables only in the manuscript, not on the local one-page abstract). Boundary conditions: benchmarks use three-dimensional periodic boundary conditions (PBC) where the underlying reference structures are bulk-like. Ensemble / thermostat / timestep (fs): per-case NVT or NVE segments with documented thermostat settings and timestep appear in the full text—N/A to list from the ingested abstract alone. Equilibration and production durations (ps/ns): likewise N/A from the abstract; see 2015dittner-venue-efficient-global. Target temperature (K): N/A on the abstract page; reproduction requires the version-of-record. Barostat / NPT: N/A — optimization is not framed as hydrostatic pressure control. Hydrostatic pressure: N/A — fitness benchmarks are not described as GPa-resolved NPT scans on the abstract page. Electric field / metadynamics / replica exchange: N/A for the global search workflow itself.

Findings

The abstract claims that the implementation delivers ReaxFF parameter sets of comparable quality with less human effort and shorter wall time than prior practice, with strong parallel scaling as the number of fitness evaluations grows. Workflow: evolutionary / genetic-algorithm search explores parameter space more globally than serial multistart local optimization. Sensitivity: performance depends on training-set construction and stochastic seeds, as for any meta-heuristic optimizer. Comparisons: claims are relative to earlier ReaxFF optimization workflows cited in the text, not new experiment. Limitations: this pdf_path is a submitted / in-press preprint; for pagination, final wording, and benchmark tables, use the version-of-record entry 2015dittner-venue-efficient-global (DOI 10.1002/jcc.23966).

Limitations

Stochastic search remains sensitive to training-set design and random seeds; global optimization does not remove chemistry judgment in selecting objectives and constraints.

Relevance to group

Software pathway for ReaxFF parameterization alongside PuReMD and LAMMPS ecosystems; canonical reference: 2015dittner-venue-efficient-global.

Citations and evidence anchors