Open Access Mechanics Datasets

Goals:

  1. Provide benchmark datasets for machine learning methods applied to mechanics problems.
  2. Provide "mechanics relevant" examples for students getting started with machine learning.
  3. Make data-driven problems in mechanics more accessible to the broad research community.

Datasets:

  1. Mechanical MNIST
    Data: https://open.bu.edu/handle/2144/39371
    Code: https://github.com/elejeune11/Mechanical-MNIST
    Manuscript: https://doi.org/10.1016/j.eml.2020.100659
    Recommended Challenge: predict scalar and full field simulation outputs from input bitmaps describing heterogeneous material properties.
  2. Buckling Instability Classification (BIC)
    Data: https://open.bu.edu/handle/2144/40085
    Code: https://github.com/elejeune11/BIC
    Manuscript: https://doi.org/10.1016/j.cad.2020.102948
    Recommended Challenge: predict if a column is "stable" or "unstable" from input vectors describing heterogeneous material properties.
  3. Right Ventricular Myocardial Mechanics (RV Mechanics)
    Data: https://dataverse.tdl.org/dataverse/RVMechanics
    Code: n/a (all experimental data)
    Manuscript: https://doi.org/10.1016/j.actbio.2020.12.006
    Recommended Challenge: define a predictive data-driven constitutive equation for fibrous soft tissue.
  4. Blood Clot Simple Shear Testing Data
    Data: https://dataverse.tdl.org/dataverse/BloodClotSimpleShear
    Code: n/a (all experimental data)
    Manuscript: https://www.sciencedirect.com/science/article/pii/S1751616120307566
    Recommended Challenge: define a predictive data-driven constitutive equation for isotropic soft tissue.
  5. 3D Printed Crossed Barrel Dataset
    Data: https://www.kablab.org/data
    Code: n/a (experimental data and ABAQUS simulations)
    Manuscript: https://advances.sciencemag.org/content/6/15/eaaz1708
    Recommended Challenge: Find the optimal crossed barrel design with the lowest number of experiments.
  6. Young and Aged Mouse Skin Biaxial Mechanical Tests
    Data: https://dataverse.tdl.org/dataverse/mouseskinmechanics1
    Code: n/a (all experimental data)
    Manuscript: https://doi.org/10.1016/j.actbio.2019.10.020
    Recommended Challenge: define a predictive data-driven constitutive equation for planar soft tissue.
  7. Asymmetric Buckling Columns (ABC) dataset
    Data: https://open.bu.edu/handle/2144/43730
    Code: https://github.com/pprachas/ABC_dataset
    Manuscript: https://arxiv.org/abs/2202.01380
    Recommended Challenge: predict the buckling direction (left vs. right) from a spatially heterogenous column geometry.
  8. Crack Propagation in Steel Structures Under Ultra-Low Cycle Fatigue
    Data: https://www.designsafe-ci.org/data/browser/public/designsafe.storage.published/PRJ-3394
    Code: https://www.designsafe-ci.org/data/browser/public/designsafe.storage.published/PRJ-3394
    Manuscript: https://stacks.stanford.edu/file/druid:ft563mt0968/Dissertation-augmented.pdf
    Recommended Challenge: predict output QoIs for the "blind prediction" experimental dataset.
  9. Ductile Fracture in Structural Steel Inclined Notch Specimens
    Data: https://purl.stanford.edu/qy227tf3022
    Code: https://purl.stanford.edu/qy227tf3022
    Manuscript: https://stacks.stanford.edu/file/druid:qy227tf3022/TR187_Smith.pdf
    Recommended Challenge: predict force vs. displacement curves and fracture probability for held out experimental samples.
  10. Cardiac Motion Analysis Challenge
    Data: http://www.cardiacatlas.org/challenges/motion-tracking-challenge/
    Code: n/a
    Manuscript: https://doi.org/10.1016/j.media.2013.03.008
    Recommended Challenge: predict cardiac motion and strain from magnetic resonance (MR) and 3D ultrasound (3DUS) imaging datasets from a dynamic phantom and 15 healthy volunteers.
  11. Unsupervised Discovery of Interpretable Hyperelastic Constitutive Laws
    Data: https://doi.org/10.3929/ethz-b-000505693
    Code: https://euclid-code.github.io/
    Manuscript: https://doi.org/10.1016/j.cma.2021.113852
    Recommended Challenge: discover strain energy density functions from emulated DIC data created via the finite element method.
  12. Discovering Plasticity Models without Stress Data
    Data: https://doi.org/10.3929/ethz-b-000534002
    Code: https://euclid-code.github.io/
    Manuscript: https://doi.org/10.1038/s41524-022-00752-4
    Recommended Challenge: discover plastic yield surfaces and hardening laws from emulated DIC data created via the finite element method.
  13. Discrete Element Traction-Separation Data
    Data: https://data.mendeley.com/datasets/n5v7hyny8n/1
    Code: https://www.poromechanics.org/software--data.html
    Manuscript: https://doi.org/10.1016/j.cma.2018.11.026
    Recommended Challenge: derive theory-consistent and microstructure-based traction–separation laws.
  14. Experimental Quasi-Static and Impact Testing of Lattices
    Data: https://www.kablab.org/lattice-quasi-static and https://www.kablab.org/lattice-impact
    Code: n/a
    Manuscript: https://doi.org/10.1016/j.matt.2022.06.051
    Recommended Challenge: leverage quasi-static tests to make predictions relevant to impact test outcomes.

Information for contributors:

  1. Consulting FAIR Principles (guidelines to improve the Findability, Accessibility, Interoperability, and Reuse of digital assets) is often helpful for effective data curation.
  2. Please make sure that the provided links will be long lasting. DOIs or handles can often be obtained through an institutional repository.
  3. Please be thoughtful about selecting licenses for data and code. For the Mechanical MNIST dataset, we chose a Creative Commons Attribution-ShareAlike 4.0 License for the data, and a MIT License for the code (we do not endorse any particular license choice).
  4. Keep file sizes small whenever possible. For example, it may make sense to create a HDF5 file.
  5. For the “Recommended Challenge,” data should be curated so that a user with limited familiarity can download all or part of the dataset and start working on the problem in just a few minutes. Think of the challenge as a starting point for what could be done with the data and code provided. Multiple datasets can have the same recommended challenge.
  6. If you have an interesting problem where regenerating the dataset from the provided code is quick and easy (i.e., can be done on a laptop in a few minutes by a user with no special expertise and with no need to install software other than Python and standard Python packages), we can link to just the code.
  7. It is not necessary to provide a link to a published manuscript. However, citing relevant manuscripts is one way to acknowledge the use of open datasets and code.

Links to other relevant pages:

  1. DesignSafe Data Depot: https://www.designsafe-ci.org/data/browser/public/
  2. NanoMine: https://materialsmine.org/wi/home
  3. NIST Materials Data Facility: https://www.materialsdatafacility.org/

Contact: E. Lejeune

If you are interested in listing your dataset, please get in touch!