lamm-mit/graphene-agent-data

Models building models for discovery of graphene metamaterial design principles in the context of failure

Companion data for the GitHub repository lamm-mit/graphene-agent and the manuscript Models building models for discovery of graphene metamaterial design principles in the context of failure. The repository holds the code, the experiment database (all 132 records with structures and stress–strain curves), the report, the figures and the app; this dataset holds the files that are too large for a code repository.

Contents

folder	files	size	description
trajectories/	132 .npz	309 MB	one file per simulation, named by run id (run_YYYYMMDD_HHMMSS_xxxxxx.npz), matching experiments/database/runs/<run_id>/record.json in the repository
large_svg_figures/	47 .svg	396 MB	the SVG versions of the atom-resolved fracture panels (each > 3 MB); paths mirror the repository (carbon_discovery/figures/progression/…, paper_analysis/figures/…)
movies/	10 + 10 .mp4, 10 .png	see movies/README.md	synchronised multi-panel fracture movies of the key trajectories (2560×1440 masters, *_1080p.mp4 for supplementary information) with poster frames and captions; no text other than the chart axes
movies/social/	28 + 28 .mp4, 28 .png	see movies/social/README.md	clean single-structure movies (14 designs × 2 fields: energy change per atom, maximum principal stress per atom), square 1440×1440 and 1080×1080, dark background, no text
SHA256SUMS.txt			checksums of every file

All simulations are athermal quasi-static uniaxial tensile tests of periodic graphene-derived sheets with the screened second-generation REBO potential (Brenner 2002; Pastewka 2008, 2013), run with the PyTorch implementation validated in the repository. Units: eV, Å; 2D stress in eV/Ų (multiply by 16.0218 for N/m).

Movies

The multi-panel movies show several designs advancing at the same engineering strain, coloured by the energy change per atom, above the stress–strain curves with moving markers. The single-structure movies in movies/social/ show one design filling the frame on a dark background, coloured by the energy change per atom (_energy) or by the per-atom maximum principal 2D stress (_stress), with freshly broken bonds flashing white. In both series the frames between the stored quasi-static states are interpolated (positions, cell, field; bonds recomputed) and are not simulated states, and the strain rate is slowed 3× around first damage, peak load and major propagation. Captions and the design/run id of every panel are in the captions.json and README.md files next to the movies; the generator is paper_analysis/make_movies.py in the repository.

Trajectory format

Each .npz contains, per stored frame of the loading sequence: positions (frames × atoms × 3, Å), cells (frames × 3 × 3), peratom_energy (eV), peratom_virial (eV, three diagonal components), coordination, eps_x, eps_y (engineering strains), sigma_xx, sigma_yy, sigma_xy (2D stress, eV/Ų), energy (eV), n_bonds, n_broken_cum (bonds counted as broken beyond 2.0 Å), and bonds_<k> (bond pairs of frame k). The reader used by all analysis scripts is carbon_discovery/analysis/fracture_viz.load_traj in the repository.

Use with the repository

git clone https://github.com/lamm-mit/graphene-agent
cd graphene-agent && pip install huggingface_hub
python scripts/fetch_data_from_huggingface.py --what all

License and citation

Please cite the manuscript above and the potential (Brenner et al., J. Phys.: Condens. Matter 14, 783 (2002); Pastewka et al., Phys. Rev. B 78, 161402(R) (2008); 87, 205410 (2013)). The simulations were carried out by an AI agent (Claude Fable 5.1 in Claude Code) as described in the manuscript.