Benchmark dataset: Phase Field fracture modeling in hyperelastic material with multiple cracks

Hyperelastic multi-crack response studied under cyclic loading simulated with FEniCS

This dataset contains 1000 finite-element simulations of a 2D square specimen with a circular cavity and eight pre-cracks, subjected to tension, unloading, and compression using a large-deformation Neo-Hookean phase-field formulation.

---

Full Dataset

Our complete collection comprises 1000 distinct Gaussian-random-field-driven boundary-condition realizations. The full dataset is available in JHU archive at: https://archive.data.jhu.edu/dataset.xhtml?persistentId=doi:10.7281/T1XFF19O

---

Visualizations

Below are first 3 samples with displacement and phase-field at the last timestep (figures), along with displacement and phase-field Evolution (animation):

Displacement and Phase-Field at Final Step (figures)

Displacement and Phase-Field Evolution (animation)

---

Apllications of this dataset

1. Benchmarking Basic ML & DL Models

Because the displacement and phase-field outputs are mapped to a uniform grid, this dataset is ideal for standard image-based architectures:

• CNN Regression

  • Map boundary-condition fields (e.g. GRF displacement maps) to phase-field damage snapshots at a given timestep.
  • Compare network depths, receptive fields, and training times to establish baseline performance (pixel-wise MSE, crack-region IoU).

• U-Net for Temporal Evolution

  • Formulate crack propagation as a sequence-prediction or segmentation task: predict the next phase-field frame from the current/previous few frames.
  • Exploit skip-connections to preserve fine crack-tip features while modeling large-scale damage patterns.

• Transfer Learning & Augmentation

  • Fine-tune pretrained vision backbones on these fracture snapshots to accelerate convergence.
  • Apply rotations, flips, and local noise to augment the 900 “train” samples and evaluate robustness.

2. Scientific Machine Learning & AI4Science

This dataset provides a playground for operator-learning and physics-informed methods:

• Neural Operators (e.g., FNO, DeepONet)

  • Learn mappings from initial displacement function to time-series of phase-field and displacement in one shot.
  • Leverage the uniform grid for efficient Fourier or kernel-based layers.

• Physics-Informed Neural Operators

  • Incorporate PDE residuals (elastic equilibrium + phase-field evolution) as soft constraints alongside data losses.
  • Use the provided energy fields to regularize models—e.g. enforce thermodynamic consistency of predicted energy dissipation.
  • Measure improvements in sample efficiency and generalization over purely data-driven or purely physics-driven approaches.

• Hybrid FE–NN Surrogate Models

  • Fuse the provided FEM solver with a neural network backbone to build FE–NN surrogates that dramatically speed up forward simulations while retaining high physical fidelity.

3. Inverse Modeling

Thanks to the elastic and fracture energy values included for the first 96 samples, this dataset supports advanced inverse analyses:

• Material Parameter Estimation

  • Infer fracture toughness (G_c), regularization length (\varepsilon), or Young’s modulus by mapping observed energy distributions (plus damage maps) back to scalar model parameters.

• Load Reconstruction

  • Given measured elastic-energy fields (e.g., from digital image correlation) and final crack geometries, invert for the most likely boundary-condition field that produced them.

---

Cite our dataset

If you use this dataset in your work, please cite it as follows:

@data{T1XFF19O_2025,
author = {Hakimzadeh, Maryam and Graham-Brady, Lori and Goswami, Somdatta},
publisher = {Johns Hopkins Research Data Repository},
title = {{Data and code associated with: Phase-Field Fracture Simulation Dataset: Hyperelastic Multi-Crack Response Under Loading and Unloading}},
year = {2025},
version = {V2},
doi = {10.7281/T1XFF19O},
url = {https://doi.org/10.7281/T1XFF19O}
}

---

Contact

In case you need more information, please feel free to contact Dr. Maryam Hakimzadeh (@mhakimz1@jhu.edu), Prof. Lori Graham Brady (lori@jhu.edu), or Prof. Somdatta Goswami (somdatta@jhu.edu).