Finite-element input histories, spatial coordinates, and output fields for the 3D lug and dog-bone cases in Sequential deep operator neural networks for plastic and thermo-viscoplastic transient material behavior, International Journal of Plasticity 201 (2026), 104689.

Paper: https://doi.org/10.1016/j.ijplas.2026.104689
Companion code: https://github.com/ncsa/T-SNOT
The same data support both S-NOT and S-DON.

This release contains the case pools selected by the companion training scripts: the first 1,000 lug cases and first 4,000 dog-bone cases, in their original order. These counts include both training and test cases. This is the two-case plasticity release; it does not include the paper's solidification benchmark. The NumPy dtypes and finite-element field values are preserved.

Data layout and units

Array axes are (case, time, node) for both output fields on disk. The input histories use 101 samples from 0 to 1 s. All .npz files contain one member, available as np.load(path)["a"]. They are NumPy arrays, not tabular datasets; download the files and load them with NumPy.

File	Shape	Meaning
`lug3d/data_amp_5000.npy`	`(1000, 101)`	Dimensionless pressure-amplitude histories; pressure equals amplitude times 100 MPa
`lug3d/xyz_coords.npy`	`(29852, 3)`	Node coordinates in metres
`lug3d/ystress_all.npy`	`(1000, 101, 29852)`	Nodal von Mises stress in Pa; the historical filename does not denote the y stress component
`lug3d/peeq_all.npy`	`(1000, 101, 29852)`	Equivalent plastic strain (PEEQ), dimensionless
`lug3d/node_labels_sorted.npy`	`(29852,)`	Abaqus node labels in the coordinate/field order; optional for training
`dogbone/Amp.npz`	`(4000, 101)`	Prescribed x displacement histories in mm
`dogbone/Coords.npy`	`(3060, 2)`	Node coordinates in mm
`dogbone/Stress.npz`	`(4000, 40, 3060)`	Nodal von Mises stress in MPa
`dogbone/PEEQ.npz`	`(4000, 40, 3060)`	Equivalent plastic strain (PEEQ), dimensionless
`dogbone/TrainIndex.npy`	`(3200,)`	Original ordered training indices within the 4,000-case pool

The filename data_amp_5000.npy is retained for compatibility, although this release contains only the 1,000 input histories that have matching outputs. The upstream input pool contained 5,000 lug histories. The upstream dog-bone pool contained 9,903 successful cases; rows beyond 3,999 are omitted here.

The lug arrays contain the initial frame plus 100 subsequent output frames. Both training scripts remove frame 0 and predict the remaining 100 frames. The dog-bone arrays already omit the initial frame and contain 40 output frames. The nominal output schedules are 0.01, ..., 1 s for lug and 0.025, ..., 1 s for dog-bone, based on the finite-element output interval settings. Exact per-frame timestamps are not stored separately in these array files.

Splits and preprocessing

Lug: the code selects 800 training and 200 test cases with seed 2024. S-DON trains once on this split; test error does not trigger a retry.
Dog-bone: preserve TrainIndex.npy exactly. The 800 test indices are the sorted complement in range(4000). Both models use this same split.

The archive preserves the extracted fields. No training normalization or additional clipping is applied during packaging. In the upstream lug extraction, PEEQ contributions were clipped to [0, 0.05] before nodal averaging. The dog-bone training scripts cap stress at 260 MPa and PEEQ at 0.6 times the maximum over the training cases. S-NOT uses standardization, and S-DON uses per-node MinMax scaling. All four scripts fit output scaling on training cases only, after splitting and any clipping, and apply the same fitted transforms to both splits. The lug training scripts convert stress from Pa to MPa.

This release provides the existing train/test split workflow. Reproducing the paper's full five-fold evaluation is outside this release's scope. These files are the simulation inputs and reference targets, not trained-model predictions or checkpoints.

Companion-code changes and evaluation limits

The companion code corrects two aspects of the source training protocol:

Output scaling and the dog-bone PEEQ cap use training cases only. The source scripts used the selected train-plus-test pool for both calculations.
S-DON lug's test-error-driven retry rule has been removed. It trains exactly once on the seeded split for the requested number of iterations.

The original first-run split rules and packaged arrays are unchanged. Synthetic regression checks verify that changing held-out targets leaves training targets unchanged, and that a large S-DON lug test error does not trigger another training call. All four model/case combinations passed small end-to-end execution checks. The corrected code has not been retrained on the full dataset to measure its effect on the paper's reported results.

The following source behaviors remain in the companion code:

Both S-NOT scripts use the test loader as validation data for checkpoint selection and learning-rate adjustment, then report errors on the same cases.
S-DON dog-bone clips saved and printed relative errors at 1.0. Its reported mean therefore differs from the unclipped mean when any errors exceed 1.0.

Consequently, S-NOT's reported test errors are not estimates from an untouched holdout used only for final evaluation. Independent evaluation would additionally require a separate validation set for model selection and a fixed test set reserved for final evaluation. The preprocessing and retry corrections do not resolve that model-selection limitation or S-DON dog-bone's error clipping. Full retraining and a revised validation/test protocol are outside this release's scope. These limitations concern the training/evaluation procedure; the packaged simulation arrays retain their original values and case order.

Download and run

After access is granted, authenticate with your own Hugging Face account and download the repository:

hf auth login
hf download jaewan-wod33/T-SNOT-data --repo-type dataset --local-dir ./benchmark-data

From the companion code repository, activate the appropriate environment and run the desired model. Replace the data path with your downloaded directory:

python -m snot.lug3d.train --data-dir /path/to/benchmark-data/lug3d --output-dir outputs/snot-lug3d
python -m sdon.lug3d.train --data-dir /path/to/benchmark-data/lug3d --output-dir outputs/sdon-lug3d
python -m snot.dogbone.train --data-dir /path/to/benchmark-data/dogbone --train-index /path/to/benchmark-data/dogbone/TrainIndex.npy --output-dir outputs/snot-dogbone
python -m sdon.dogbone.train --data-dir /path/to/benchmark-data/dogbone --train-index /path/to/benchmark-data/dogbone/TrainIndex.npy --output-dir outputs/sdon-dogbone

No training-code changes are needed for these subsets. Each default sample count matches the corresponding dataset. The coordinate arrays are shared across cases. Full training requires substantially more memory than a single array file; consult the companion code's environment and execution instructions.

Integrity and provenance

manifest.json records shapes, dtypes, file sizes, SHA256 hashes, value ranges, and finite-value checks. The cropped array payloads are checked against the corresponding original prefixes. All other array files are unchanged.

The default splits have no overlapping case indices: 800/200 for lug and 3,200/800 for dog-bone. Every released input history was also checked against its original finite-element input deck (1,000 lug and 4,000 dog-bone cases). No identical input histories or identical paired stress/PEEQ output fields were found across these train/test splits. These checks establish file integrity and case separation; they do not remove the training-protocol limitations above.

Verify downloaded files from the dataset directory with:

sha256sum -c SHA256SUMS

The release excludes finite-element ODB files, model weights, prediction results, noise-augmented arrays, and simulation runtime logs. Access is managed by the dataset owner. The companion code's MIT license is not a declaration of the license for these data.

Citation

@article{park2026sequential,
  title = {Sequential deep operator neural networks for plastic and thermo-viscoplastic transient material behavior},
  author = {Park, Jaewan and Kobayashi, Kazuma and Liu, Qibang and Koric, Amar Alem and Abueidda, Diab W. and Alam, Syed Bahauddin and Koric, Seid},
  journal = {International Journal of Plasticity},
  volume = {201},
  pages = {104689},
  year = {2026},
  doi = {10.1016/j.ijplas.2026.104689}
}

Downloads last month: 13

Total file size:

50.7 GB