Skip to content

Optimization of ReaxFF Reactive Force Field Parameters for Cu/Si/O Systems via Neural Network Inversion with Application to Copper Oxide Interaction with Silicon

Galley / SI sibling

Corpus pdf_path is a galley (Roshan_JPCC_CuSiO_2023_galley.pdf). The catalog also lists an SI-focused ingest under 2023roshan-venue-paper; align figure/table citations with the VOR when available. See docs/corpus/NON_PRIMARY_ARTICLE_PAPER_SLUGS.md section A.

Summary

Roshan et al. introduce a deep-learning-accelerated ReaxFF optimization workflow in which a neural network learns to predict ReaxFF observables from force-field parameters, enabling neural inversion searches that adjust parameters to match DFT and experimental references more efficiently than brute-force scans. The target application is a Cu/Si/O ReaxFF relevant to microelectronics processing contexts where copper and copper oxide species interact with silicon substrates and interfaces. After optimization, the authors run large reactive MD cells (reporting systems up to ~3542 atoms) to examine diffusion and interaction patterns for Cu oxides adjacent to Si, emphasizing temperature-dependent transport and the role of oxygen in promoting Cu ingress.

Methods

QM training data and ReaxFF objectives (A)

  • DFT references for Cu/Si/O stoichiometries and configurations (oxides, interfaces, clusters—see article tables).
  • Fitting targets include heats of formation, relative phase stability, and additional energy/structure metrics enumerated in JPCC Methods/SI.

Neural-network–accelerated optimization (A)

  • Surrogate: A neural network maps ReaxFF parameters to predicted observables, enabling inversion / search with lower wall-clock cost than brute-force scans (comparison protocol and stopping criteria in the paper).
  • Software: LAMMPS-compatible ReaxFF parameter sets; optimization workflow details (hyperparameters, training losses for the NN) are specified in the manuscript.

Reactive molecular dynamics applications (B)

  • Systems: Large cells up to ~3542 atoms (per abstract) containing Si substrates and Cu / copper oxide species.
  • Conditions: Temperature-dependent diffusion and interaction studies comparing Cu oxide vs metallic Cu clusters.
  • Analysis: Transport and penetration trends of Cu and O near Si interfaces.

MD application (Cu/Si/O after fit)

Engine / code: LAMMPS with the optimized Cu/Si/O ReaxFF (up to ~3542-atom cells in the abstract). Si substrate and Cu / copper oxide species in 3D PBC; T-dependent diffusivity/transport; timestep, equilibration/production (ps/ns), ensemble, thermostat/QEq in the JPCC text (this galley path: confirm VOR). N/A — no static interfacial electric field; N/A — no metadynamics/replica sampling beyond the reported RMD; N/A for constant-volume work, no NPT barostat unless the VOR specifies NPT with hydrostatic pressure control.

Findings

Optimization performance

Neural inversion achieves shorter average optimization time than brute-force search while improving or maintaining agreement with training-set metrics such as heats of formation and relative phase stability (quantitative tables in JPCC).

Simulations report strong temperature dependence for Cu oxide mobility; Cu oxide aggregates can diffuse faster than metallic Cu clusters adjacent to Si under the authors’ conditions.

Role of oxygen

Oxygen content strongly modulates Cu ingress into silicon in their trajectories—exact diffusivities and barriers are listed in the article/SI.

Limitations

Galley PDF in corpus; verify final pagination and SI tables against VOR. Machine-learning acceleration does not remove physics limitations of ReaxFF itself. Microelectronics-adjacent claims should be checked against the training scope (defect types, oxidation states, and temperature windows).

Reader notes (navigation)

Cross-link responsibly with 2023roshan-venue-paper (SI-focused ingest) when discussing supporting tables; keep DOI 10.1021/acs.jpcc.3c03079 as the stable anchor for automation.

Relevance to group

Core contribution on ML-accelerated ReaxFF fitting for Cu/Si/O chemistry with van Duin authorship.

Citations and evidence anchors