cr15pykrnl/knet2-auxetic-deformation-partial

KNet2 auxetic deformation: partial recovery

Version 0.1.1. Public dataset: cr15pykrnl/knet2-auxetic-deformation-partial. This release preserves available simulation-derived deformation images and associated records from the separate KNet2 neural-network extension draft. It is incomplete. It is not the published KNet1 paper's reproduction dataset, not a verified benchmark, and contains no checkpoint or accuracy claim.

Source and relationship to the paper

The recovered draft is titled KNet2: Multi-layer Neural Network and k-means Clustering for Detecting Deformation Mechanisms in Re-entrant Honeycomb Auxetics, by Gaurav Singh, Rahul Singh Dhari and Zia Javanbakht. No publication identifier has been established for this draft.

Prior work: Automated detection of deformation mechanisms in re-entrant honeycomb auxetics using machine learning, International Journal of Protective Structures 16(4), 853–877 (2025). That paper uses logistic regression; its results must not be attributed to this KNet2 recovery. The original data conversion predates source Git tracking; its exporter hash remains in metadata/provenance.json.

The maintained paper companion is AuxMech. It provides optional download and inspection helpers for this related dataset; it does not claim to reproduce the published KNet1 implementation or release a complete KNet2 training pipeline. The newly authored Python helpers use MIT; this dataset retains CC BY 4.0.

Version 0.1.1 repairs the Parquet image-storage types so clients and the HF viewer recognize actual images. Every row, numeric value and embedded PNG byte is unchanged from version 0.1.0. The nullable image path now has Arrow string type rather than null type. Verification evidence is in metadata/validation.json. The data/schema release is pinned at 1b2ba3915a38410cfb3dcbafb71e41f1a7c426d0. Later README/checksum amendments create a separate card revision; use this data revision for repeatable analysis rather than relying on a moving branch.

Contents and collection

• 2,482 recovered registered grayscale images, each (126, 199) pixels (height × width), with original frame references and loading-angle values.
• Ten per-angle registration composites, including recovered bounding boxes.
• All 24 recovered K-means cluster IDs; notebook mode mapping included as an explicitly unverified annotation proposal, not expert ground truth.
• One raw video and seven raw force/displacement CSV tables, plus the original metadata and an explicit availability inventory.
• Thirty-eight recovered clustering-score rows, not rerun evaluations.

Recovered code indicates grayscale conversion, inverse thresholding, rotation, edge removal, dilation, scaling, per-angle bounding-box registration and intensity normalisation. See metadata/schema.json for the recovered settings and caveats. Raw assets were copied unchanged. Images were losslessly converted from two hash-verified recovered pandas pickle sources into PNG bytes embedded in Parquet. The two recovered image collections agree on all samples and pixels, so images are stored once. Source file hashes are in metadata/provenance.json.

Schema, labels and units

processed/images.parquet contains sample_id, theta_degrees, frame_index, recovered_stress, recovered_strain, cluster_id, notebook_mode_code, notebook_mode_id, image_sha256 and image. Angles are in degrees as indicated by recovered code/draft context; frame indices are preserved without changing their original zero-based reference convention. Images are uint8, range 0–255. Their original shape is retained; they are not resized to the network's 64 × 64 input.

Physical units of the raw dy, f and recovered stress values have not been independently established. recovered_strain is dimensionless by its intended meaning, but its calculation's unit consistency has not been independently established. Numeric values are preserved without conversion or a unit claim. Do not combine these values with external physical measurements without resolving units. All numeric output values were checked for finiteness.

The notebook mode codes are X, YY, UD, DN, DX, VV. Their integer IDs and the complete cluster-to-code mapping are in metadata/schema.json. Codes are retained as written in the source; expansions and scientific correctness are not asserted.

Splits and limitations

recovered is an inventory split, not a training, validation or test partition. No historical evaluation partition is reconstructed. Samples from one angle may share a simulation sequence, and image duplicates may cross angles or frames. There are 213 repeated image hashes beyond their first occurrence. Split by verified simulation/sequence identity and duplicate lineage before training or evaluation; angle alone is not proven to identify an independent run.

Nine videos and three force/displacement tables named by raw metadata are absent. The available video has no paired table. Thus no complete raw video/table pair is available, and the raw pipeline cannot be reproduced end to end from this release. The 2,482 derived samples cover ten angles despite missing raw assets; existence of derived samples is not proof of complete raw provenance.

Features/PCA state, original input images, the semantic-label pickle and model checkpoint are absent. Balanced data and pickled runtime dataloaders are excluded. No model loading, training, accuracy reproduction or independent annotation validation has been performed. Clustering scores lack the recovered feature matrix needed to rerun them.

Subsequent recovery audit (2026-10-24)

A reconstruction of the recovered helper's 80:20 seed-42 row split on the 6,000-row balanced collection found 574 shared source-frame keys. Of 1,200 validation rows, 1,018 have a source key also present in training. This is a reconstruction from recovered code and data, not the saved historical runtime split or a new performance measurement. The balanced collection is excluded from this release. Split source sequences and duplicate lineage before resampling.

At 22.5 degrees, the available raw force/displacement table does not reproduce the recovered image metadata under the supplied formulas. Among 488 records, 484 stress and 487 strain values differ exactly; maximum absolute differences are approximately 0.05217647 and 0.01155979 in unresolved recovered units. Six other available angles reproduce the arithmetic exactly. Both versions are preserved; resolve their source/version relationship before physical analysis. These checks establish neither physical units nor annotation correctness.

Download and loading

Download this reviewed snapshot of the data and corrected dataset card (its data files match the repaired-data revision above):

hf download cr15pykrnl/knet2-auxetic-deformation-partial \
  --repo-type dataset --revision c72a5a713168a8a65de25fa32e20bfd241d7afec \
  --local-dir data/knet2

Or load the image inventory using the HF datasets client:

from datasets import load_dataset
samples = load_dataset(
    "cr15pykrnl/knet2-auxetic-deformation-partial",
    "recovered-images", split="recovered",
    revision="1b2ba3915a38410cfb3dcbafb71e41f1a7c426d0",
)
image = samples[0]["image"]

The separate registration-composites configuration selects the ten per-angle composites. Both use the recovered inventory split, not a training/test split. No executable dataset script is required.

For a downloaded copy, the verified PyArrow/Pillow path is:

import io
import pyarrow.parquet as pq
from PIL import Image
samples = pq.read_table("data/knet2/processed/images.parquet").to_pandas()
image = Image.open(io.BytesIO(samples.iloc[0]["image"]["bytes"]))

The repaired files were checked with Python 3.12.9, NumPy 1.26.4, pandas 2.2.3, PyArrow 19.0.1, Pillow 11.3.0 and datasets 3.6.0. Local HF datasets loading recognizes the Image feature and decodes every sample/composite. This verifies file/client compatibility; no training, inference or physical validation is claimed. The live HF viewer now recognizes Image features in both configurations, and sample/composite previews were fetched and decoded. HF serves JPEG previews; use the unchanged embedded PNGs for pixel analysis. These checks do not establish scientific validity.

Licence and attribution

CC BY 4.0, selected by the rights-confirming dataset owner on 2026-10-24. See LICENSE. Attribute the original research authors above and identify this artifact as KNet2 auxetic deformation: partial recovery, version 0.1.1, maintained by cr15pykrnl. Cite the immutable data revision above. The prior paper DOI is background, not a DOI for this dataset.

For schema, omissions and verification results see metadata/; for exact release bytes see checksums.sha256. Checksums establish byte integrity, not scientific validity or completeness.