Multiobjective genetic training and uncertainty quantification of reactive force fields
Summary¶
Reactive force fields such as ReaxFF are widely used to reach nanosecond-to-microsecond reactive trajectories, but parameterization is often posed as a single-objective fit to a limited set of quantum references. This npj Computational Materials article introduces an in situ multiobjective genetic algorithm (iMOGA) that trains ReaxFF models while comparing reactive MD (RMD) trajectories to ab initio MD (QMD) on the fly, maintaining a Pareto front over multiple quantities of interest (QoI) rather than collapsing everything into one score. The workflow is engineered for scalability: coupled RMD, QMD, and GA ranks communicate through interprocess communication to avoid disk-bound trajectory exchanges that would bottleneck large-scale optimization. The demonstration application targets high-temperature sulfidation of MoO\(_3\) by H\(_2\)S, a far-from-equilibrium process relevant to MoS\(_2\) chemical vapor deposition chemistry, where trajectory-level agreement is arguably more informative than static energy matching alone.
Methods¶
Optimization uses a multiobjective genetic algorithm with nondominated sorting to retain Pareto-optimal ReaxFF parameter sets that jointly address competing QoI targets (rather than averaging errors into a single loss). Training compares RMD and QMD trajectories generated during optimization, enabling the GA to reward models that reproduce time-dependent observables of the reactive environment. The software coupling strategy (detailed in the article) distributes work across ranks and minimizes file I/O overhead. For the Mo–O–S–H demonstration, the authors refine parameters appropriate to H\(_2\)S reaction with MoO\(_3\) flakes at elevated temperature, with reference QMD settings and RMD integration choices specified in the methods and supplementary materials of the publication.
MD application (sulfidation demonstration)¶
Molecular dynamics couples reactive MD (ReaxFF) to ab initio MD (QMD) trajectories inside the optimization loop; the publication and SI name the MD driver(s) and integration settings (N/A — LAMMPS vs other engine not duplicated on this summary page). System & composition: MoO\(_3\)-derived flake supercells exposed to H\(_2\)S-relevant Mo–O–S–H chemistry (atom counts and stoichiometry in Methods). PBC: three-dimensional periodic cells for the reactive sulfidation setup. Ensemble / thermostat / barostat / timestep / duration: N/A — NVE/NVT/NPT labels, thermostat type, fs timestep, and equilibration vs production ns not transcribed here—copy from the PDF/SI before reproducing runs. Temperature: elevated thermal conditions for high-rate sulfidation (exact K in Methods). Pressure: N/A — hydrostatic pressure protocol not summarized in this abstract-level note; confirm whether NPT appears in the primary Methods. Electric field: N/A — not used. Enhanced sampling: N/A — not indicated beyond genetic exploration of force-field parameters.
Force-field training (iMOGA)¶
Parent / scope: ReaxFF for Mo–O–S–H updated in situ during iMOGA. QM reference: QMD trajectories supply DFT-level energies and forces for on-the-fly comparison. Training set: reactive H\(_2\)S + MoO\(_3\) environments emphasizing trajectory-level QoI rather than static energy points alone. Optimization: multiobjective genetic algorithm with Pareto sorting over competing QoI. Reference data: QMD benchmarks and published QM settings in Methods/SI.
Findings¶
The authors report that trajectory-level training can approach QMD behavior for the sulfidation process studied, in the sense summarized by their QoI metrics and comparative plots. A central claim is that the Pareto ensemble of force-field solutions provides uncertainty quantification (UQ) that single-objective fits do not: practitioners can inspect spread across nondominated models when making predictions. For the MoO\(_3\) sulfidation case, the work is presented as a proof that practical ReaxFF refinement for CVD-relevant reactive pathways can be carried out with explicit multiobjective structure, not only as a post hoc sensitivity check.
Corpus honesty. QoI definitions and numerical tolerances are in the PDF/SI; this summary tracks the open-access abstract language only.
Limitations¶
Cost remains high even with scalable coupling; QMD reference quality, time-step choices, and ReaxFF functional limits still bound transferability beyond the training chemistry.
Relevance to group¶
Documents a modern data-driven ReaxFF optimization paradigm with explicit UQ, complementary to manual parabolic training workflows common in hand-tuned ReaxFF development.