Skip to content

Multiparameter and Parallel Optimization of ReaxFF Reactive Force Field for Modeling the Atomic Layer Deposition of Copper

Summary

This Journal of Physical Chemistry C article develops a multiparameter, parallel optimization workflow—organized with Taguchi experimental-design ideas—to refine ReaxFF parameters for Cu/C/H/N/O chemistry aimed at atomic layer deposition (ALD) modeling. The practical target is copper metallization from organometallic precursors with atomic hydrogen as a coreactant, a regime where ligand fragmentation, surface contamination, and incomplete elimination can undermine film purity. The authors position the work as extending prior Cu ReaxFF descriptions by tightening interactions needed for ALD-relevant bond-making and bond-breaking events on Cu surfaces.

Methods

Force-field training (ReaxFF). The authors extend an existing Cu reactive description to Cu/C/H/O/N by optimizing Cu–C, Cu–N, and Cu–H (plus oxygen-containing) interaction blocks needed for copper ALD. Density functional theory (DFT) reference energies, forces, and reaction pathways supply the training set, while a Taguchi-inspired multiparameter parallel optimization workflow searches parameter space efficiently before validation on organometallic motifs. QM details (functional, basis set, k-point sampling) belong to the article Methods and should be read from papers/ReaxFF_others/Hu_Copper_JPCC_2017.pdf.

Molecular dynamics (reactive). After the fit, reactive molecular dynamics (RMD) simulations implement abbreviated ALD cycles on Cu for [Cu(iPr-amd)]\(_2\) versus Cu(dmap)\(_2\) with H-radical coreactant pulses, monitoring adsorption, fragmentation, and residue populations at temperature setpoints (K) and thermal coupling recipes tabulated in the JPCC Methods. The indexed abstract does not restate supercell sizes, exact atom counts, timestep (fs), thermostat/barostat choices, NVT/NPT labels, or equilibration/production duration (ps/ns); extract those from the JPCC PDF rather than this wiki summary. Periodic models are implied for metallic slabs with periodic images. Electric fields and metadynamics/umbrella enhanced sampling are not highlighted in the excerpted pages.

Static QM / DFT. DFT supplies the training references; it is not used as the production AIMD engine for large-cycle RMD.

Review scope. N/A — primary parameterization plus demonstration trajectories.

Findings

Outcomes and mechanisms. For [Cu(iPr-amd)]\(_2\), the first half-cycle favors dissociative adsorption into Cu(iPr-amd)-derived surface species; the H-radical pulse eliminates some fragments but leaves nitrogen-containing residues such as C\(_5\)H\(_{12}\)N\(_2\) and C\(_2\)H\(_4\)N that can contaminate the interface. For Cu(dmap)\(_2\), adsorption splits into Cu(dmap) plus dmap, and a subsequent hydrogen-transfer sequence drives dmap ligands into the gas phase more completely.

Comparisons. Versus the amidinate route, Cu(dmap)\(_2\) exposes simpler surface reaction sequences with fewer persistent C–N fragments in the abbreviated-cycle model, aligning with the abstract’s “cleaner ALD chemistry” message.

Sensitivity / design levers. Precursor identity (amidinate versus alkoxide) is the dominant lever in the modeled abbreviated cycles; H-radical exposure controls how completely ligands are stripped versus trapped as contamination.

Limitations / outlook. Abbreviated cycles omit full industrial pulse/purge timing; ReaxFF barriers remain empirical relative to experiment.

Corpus honesty. Protocol numerics beyond the abstract must be taken from the PDF at pdf_path; the local extract stops early in the introduction.

Limitations

Abbreviated cycles omit full industrial pulse/purge timing and chamber chemistry; contamination predictions are force-field dependent.

Reproducibility notes

For LAMMPS-style workflows, operators should archive the distributed ReaxFF parameter file used with the publication, the element typing rules for organometallic ligands, and the H-radical introduction protocol (flux, coverage, and thermostat coupling), because reactive ALD outcomes can be sensitive to radical flux approximations. When extending the model to new precursors, re-run the Taguchi optimization stage rather than borrowing parameters piecemeal from unrelated Cu datasets.

The Taguchi framing is not decorative: it is intended to reduce the number of explicit multiparameter trials while exploring coupled corrections to Cu–C, Cu–N, and Cu–H terms that would otherwise explode combinatorially. Reproducing the optimization therefore requires logging which parameter groups were varied together, the levels tested, and the objective function used to score training-data agreement—details that belong in the article/SI rather than being inferred from a single final parameter file.

Relevance to group

Demonstrates modern ReaxFF optimization workflows (parallel/Taguchi) for metallization chemistry relevant to interconnect processing.

Citations and evidence anchors