Skip to content

Multiple Objective NSGA-II-Based Optimization Program and Its Application in Reactive Force Field for 2,4,6-Trinitrotoluene Diffusion in the Aqueous Phase

Summary

2,4,6-trinitrotoluene (TNT) environmental chemistry couples solid-state packing with aqueous reaction pathways that are difficult to probe atomistically without reactive potentials. Zhang, Li, and Liu introduce MONOP, a multi-objective genetic optimizer patterned on NSGA-II, to retrain ReaxFF parameters for TNT in liquid water beginning from the community CHON-2017_weak parent set. The objective landscape mixes crystallographic observables for TNT with liquid-structure metrics so the optimizer cannot sacrifice lattice fidelity to fix water alone. A parallel mean-square displacement (MSD) analysis compares TNT translational diffusion to hydroxide mobility to interrogate whether hydrolysis is limited by bulk transport or interfacial attack.

Methods

MONOP evolves populations of ReaxFF parameter vectors under Pareto selection, updating generations until multi-objective tolerances are met—the abstract cites convergence near 11 generations for this application. Each candidate undergoes TNT crystal simulations to extract lattice constants (a, b, c), mass density, and element–element radial distribution functions, stacked against the same benchmarks for CHON-2017_weak and ReaxFF-lg references. Liquid-phase cells evaluate water density and transport statistics. MSD curves for TNT and OH⁻ are accumulated from NVT-style production segments at states specified in the article, enabling qualitative rate comparisons without invoking full reactive rare-event sampling for every channel.

Force-field training. MONOP is a multi-objective NSGA-II-style genetic algorithm optimizer applied to ReaxFF parameters for TNT/water, starting from CHON-2017_weak as the parent parameter set. QM/DFT training references and experiment targets for crystal lattice metrics and liquid water density are defined in the article and Supporting Information; parameter optimization balances those objectives on the Pareto front rather than a single least-squares score.

MD application (article §2–3). Engine / code: ReaxFF molecular dynamics driven through MATLAB-orchestrated parallel LAMMPS in the MONOP workflow. Ensemble / stages: NVT and NPT MD as labeled in the Methods; for all MD runs, timestep 0.25 fs; Berendsen thermostat; the Methods line after the 0.25 fs timestep gives paired damping constants 500 fs and 10 fs for NVT vs NPT MD segments (read the PDF wording—this summary does not reassign which constant applies to pressure vs temperature). RDF benchmarks: TNT crystal runs of ~25 ps total time; RDF averaged every 50 timesteps over the last 20 ps with lattice constants and density from that window. Liquid box: NPT relax ~25 ps, then ~75 ps production for water RDF pairs. MSD: after NPT equilibration of the aqueous cell, NVT MSD analysis compares TNT and OH⁻ migration (see Results / Figure 6). System / PBC: PBC TNT crystal and TNT/water liquid supercells with atom totals given in the SI/Methods tables; ~100 K crystal RDF verification reduces thermal noise, while room-temperature aqueous cells feed liquid metrics and MSD analysis. Electric field / enhanced sampling: N/A — NSGA-II is the outer search, not metadynamics.

Findings

MONOP reaches competitive Pareto fronts quickly, indicating that NSGA-II-style search pairs well with ReaxFF’s moderate-dimensional parameter tweaks when observables are well posed. The optimized TNT/water parameterization reproduces TNT crystal metrics closely while leaving liquid water density modestly high relative to experiment, a common stress test for CHON water subsets. MSD analysis reports TNT diffusion orders of magnitude slower than OH⁻ transport, supporting a mechanistic picture where interfacial hydroxide attack gates hydrolysis rather than bulk TNT migration. Operators should cross-check barriers for specific reaction channels with QM because genetic fits can overfit tabulated observables. The Xi’an Jiaotong authorship line situates the work as an external consumer of CHON-2017_weak parameters; reproducibility requires archiving the exact MONOP parameter files alongside the published force-field tables.

Limitations

External Monte Carlo / genetic optimization can overfit training observables; validate reaction pathways independently with QM. Xi'an Jiaotong authorship—method is adjacent to, but not authored inside, the Penn State CHON-2017_weak line.

Relevance to group

Demonstrates downstream reuse of CHON-2017_weak for energetic-material aqueous chemistry and automated multi-objective ReaxFF tuning.

Citations and evidence anchors