Skip to content

INDEEDopt: a deep learning-based ReaxFF parameterization framework (uncorrected proof PDF)

Corpus PDF role

Uncorrected proof PDF (Sengul_npj_CompMat_2021_INDEED_opt_galley.pdf). The typeset journal PDF is Sengul_npj_CompMat_2021_INDEED_opt.pdf on 2021sengul-npj-computat-indeedopt-deep.

Summary

Modern ReaxFF projects routinely involve hundreds of bonded and off-diagonal parameters, so global search strategies that combine design of experiments with surrogate models are increasingly attractive for group-scale fitting pipelines. ReaxFF optimization is high-dimensional and often non-convex: conventional searches can stall in poor local minima or waste effort exploring unphysical regions of parameter space. INDEEDopt (INitial-DEsign Enhanced Deep learning-based OPTimization) combines Latin Hypercube Design (LHD) sampling with a deep learning model that learns error structure across sampled parameter vectors. The workflow first explores broadly, then uses the learned surrogate to prune unpromising domains and focus refinement on low-error basins. The article demonstrates the approach on reparameterizing a nickel–chromium alloy ReaxFF and a tungsten–sulfur–carbon–oxygen–hydrogen (quinary) system, comparing wall time and training-error metrics against conventional optimizers under comparable QM training sets.

Methods

Corpus PDF. This page tracks the uncorrected proof PDF; authoritative tables, SI pointers, and typeset text are on [[2021sengul-npj-computat-indeedopt-deep]].

1 — MD application. The paper centers on ReaxFF parameter fit; any short MD in LAMMPS to validate a field (see the joint article) is NVT molecular dynamics in PBC supercells with a thermostat, timestep in fs, and ps/ns segments as stated there, temperature in K, and N/A / optional NPT barostat for the cited validation snippets. N/A — external electric field in the INDEEDopt workflow; N/A — umbrella as the headline sampling (search is in parameter space).

2 — Force-field training (INDEEDopt). As on the VOR: LHD over ReaxFF parameter vectors, neural surrogate to rank low error manifolds, and conventional optimizer baselines on the same QM reference / validation data for Ni–Cr and W–S–C–O–H demonstrations; ReaxFF equation (bond/angle/…) parrex-style optimization and DFT reference energies in the training set are documented on [[2021sengul-npj-computat-indeedopt-deep]].

3 — Static QM / experiments. QM reference data: see VOR. N/A — new laboratory experiment in the methods paper.

Findings

In the reported Ni–Cr and W–S–C–O–H tests, INDEEDopt identifies parameter sets with lower training error faster than the conventional route, reflecting reduced wasted effort in non-viable regions of parameter space. The LHD → deep learning pipeline is presented as a practical way to escape poor local minima when initial guesses are weak, complementing (rather than replacing) domain expertise in training-set curation.

Comparisons and corpus honesty. All quantitative comparisons to conventional ReaxFF optimization and DFT errors should be read from the typeset VOR on [[2021sengul-npj-computat-indeedopt-deep]]; this proof PDF is not a substitute for the version-of-record.

Limitations

Performance depends on training-set coverage, network architecture, and LHD extent; transfer to new element subsets requires fresh sampling and validation. Proof-PDF figure/page alignment may differ slightly from the VOR on 2021sengul-npj-computat-indeedopt-deep. As with any accelerated optimizer, reported gains are demonstration-dependent: practitioners should still verify physics of resulting parameters on hold-out QM data and independent MD benchmarks before production use. Repository automation maps this stable paper_id to normalized/papers/2021sengul-npj-computat-indeedopt-deep-2.json and the repo-relative pdf_path. Where extraction_quality is partial, the tracked PDF and DOI remain the quantitative authority over short local extracts.

Reader notes (navigation)