Published 2026 | Version v1

Dataset for the publication "Data-efficient Bayesian-guided design selection from large candidate sets: Application to hyperelastic stochastic metamaterials"

  • 1. ROR icon Technische Universität Braunschweig

Description

Datasets and cached computational results supporting the paper:

 

Data-efficient Bayesian-guided design selection from large candidate sets: Application to hyperelastic stochastic metamaterials", Materials & Design (2026), 116984, DOI: https://doi.org/10.1016/j.matdes.2026.116984

 

The accompanying source code is available at https://github.com/hooman-danesh/BayGDS.

 

CONTENTS

A single archive, baygds-data.tar.gz.

 

FORMAT

Arrays are NumPy .npz, model checkpoints are PyTorch .pt, and configurations and provenance records are JSON. Every array, its shape, and its units are documented in docs/DATA.md in the source repository.

 

Raw inputs

data/structures/structures.npz 50,000 binary metamaterial unit cells on a 96 x 96 periodic grid.
data/pca/pc_scores.npz Principal components of the solid and interface two-point correlations (50,000 x 8).

data/oracle/oracle_0deg.npz

data/oracle/oracle_45deg.npz

High-fidelity effective first Piola-Kirchhoff stress responses in MPa for all 50,000 structurzes over 105 deformation states, at 0 and 45 degree loading orientations, together with the applied deformation gradients and loading-path tags. The 0 degree set trains the surrogate; the 45 degree set defines the held-out design targets.

 

Cached results

output/surrogate/active_learning/ The multi-output variational Gaussian-process surrogate selected by the active learning stopping rule at 200 observed structures, with its training configuration, normalization statistics, and the full acquisition trajectory.
output/inverse_design/ Cached design selection results for 1,000 held-out targets across seven target stress-component combinations and 21 oracle budgets.
output/studies/pc_dimension/ Surrogate accuracy against the number of retained PC scores.
output/studies/lambda_sensitivity/ Design-selection performance against the uncertainty weight.
output/studies/baseline_comparison/ One directory per search method: proposed, random_search, bo_ei_engineered_features, and bo_ei_raw_image_pcs (BO-EI on six principal components of the raw binary images).

 

Disclaimer

This dataset is provided "as is" without any express or implied warranties. The user is solely responsible for any consequences resulting from its use.

 

Dataset Origin

The topologies of the metamaterial unit cells used in this work are based on the publicly available dataset from:

J.H. Bastek & D.M. Kochmann, Inverse-design of nonlinear mechanical metamaterials via video denoising diffusion models: dataset and model checkpoints. ETHZ Research Collection (2023). https://doi.org/10.3929/ethz-b-000629716 

J.H. Bastek & D.M. Kochmann, Inverse design of nonlinear mechanical metamaterials via video denoising diffusion models. Nat Mach Intell 5, 1466–1475 (2023). https://doi.org/10.1038/s42256-023-00762-x

The original dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

 

CITATION

If you use these data, please cite the paper:

H. Danesh and H. Wessels, Data-efficient Bayesian-guided design selection from large candidate sets: Application to hyperelastic stochastic metamaterials, Materials & Design (2026), 116984. DOI: https://doi.org/10.1016/j.matdes.2026.116984

Files

Files (266.7 MB)

Name Size
md5:ba82d4596d7758a7c5d8d8f0e6cd16e7
266.7 MB Download

Additional details

Software

Repository URL
https://github.com/hooman-danesh/BayGDS
Programming language
Python
Development Status
Active

References