Skip to content

A Graph Dynamical neural network approach for decoding dynamical states in ferroelectrics (galley PDF)

Summary

This slug records a galley or corrected proof PDF byte-identical ingest for the Carbon Trends article with DOI 10.1016/j.cartre.2023.100264. Galley files often differ from the final journal XML in line breaks, figure resolution, and occasionally wording; for authoritative statements, readers should use the version-of-record wiki page [[2023dhakane-carbon-trend-graph-dynamical]], which points at the non-galley PDF path in this corpus when available. Scientifically, the publication combines large-scale ReaxFF molecular dynamics of pristine and oxygen-vacancy-containing barium titanate with graph-based dynamical neural networks and Markov-state or Koopman-style analyses on local polarization features, aiming to separate slow and fast relaxation modes at domain walls and defects. The work sits at the intersection of reactive ferroelectric MD, machine learning, and reduced dynamical models for complex oxide interfaces. Ferroelectric domain walls are inherently heterogeneous; combining reactive trajectories with graph dynamical models is one route to compress wall motion into interpretable slow modes for device-scale reasoning.

Methods

Corpus role (duplicate ingest / non-primary PDF)

This slug tracks papers/Dhakane_Ganesh_CarbonTrends_GraphNN_BaTiO3_2023_galley.pdf bytes separately from the primary ingest for hash and layout provenance. It is not the canonical Methods source.

Where the protocols live

LAMMPS system sizes (~10⁴–10⁵ atoms), PyTorch graph model definitions, thermostat/timestep choices, and local polarization featurization are documented on [[2023dhakane-carbon-trend-graph-dynamical]] and the peer-reviewed PDF for DOI 10.1016/j.cartre.2023.100264.

For retrieval-aligned detail (mirroring the VOR page): reactive trajectories use ReaxFF for BaTiO₃ with oxygen vacancies, building on the authors’ prior PCCP 2019 parametrization; cells reach roughly tens of thousands of atoms with pristine and defective domain-wall setups. Local polarization vectors distinguish bulk-like, wall, and defect-adjacent environments, then feed graph dynamical networks and Koopman/Markov analyses implemented in PyTorch to separate fast vs slow collective modes.

MD application (N/A for duplicate slug)

N/A — this galley duplicate is not the canonical Methods source. MD protocol (engine, LAMMPS ReaxFF settings, PBC, NVT/NPT, temperature setpoints (e.g. 300 K if used in the VOR), ps/ns run lengths, barostat, field/wall strain, enhanced sampling): use [[2023dhakane-carbon-trend-graph-dynamical]] and the version-of-record PDF for DOI 10.1016/j.cartre.2023.100264.

Findings

Substance mirrored from the VOR page

Oxygen vacancies create localized slow dipole relaxation and defect dipoles that impede domain-wall motion; wall segments can differ by ~order-of-magnitude effective dynamics in the ReaxFF + graph-dynamical analysis. The VOR discussion also highlights spatial heterogeneity along rough walls and cites ~10× spreads between fast wall segments and slow, high-curvature segments relative to mean wall behavior—use [[2023dhakane-carbon-trend-graph-dynamical]] for the full quantitative discussion.

What to cite for science

Use [[2023dhakane-carbon-trend-graph-dynamical]] (or the VOR PDF) for figures, numbers, and quantitative claims—this galley page exists for manifest alignment and deduplication by DOI. Corpus honesty: duplicate PDF; sensitivity to temperature/field/strain is not re-derived on this page.

Limitations

Galley pagination does not match the final article; cite the DOI and primary wiki slug for quotations. ReaxFF accuracy for ferroelectric oxides is context-dependent and should be cross-checked against experiment for quantitative barriers. Downstream chunk builders should key embeddings to the VOR markdown body, not duplicate galley boilerplate. Graph-neural sections on the primary page list tensor shapes omitted here.

Relevance to group

Duplicate ingest path for Dhakane–Ganesh ReaxFF plus ML ferroelectric work with van Duin-network co-authorship.

Citations and evidence anchors

Reader notes (navigation)