A benchmark for scientific machine learning

FRACTURE-BENCH

Fast neural surrogates for phase-field fracture

Compare operator-learning architectures on two high-fidelity fracture datasets, then see how they predict crack initiation and propagation in a single forward pass.

2physics datasets
5model families
5,000FEM simulations
mssurrogate inference

The basics

Fracture-Bench brings neural surrogate models to phase-field fracture simulation. It compares operator-learning and generative architectures on shared data splits and evaluation metrics across two high-fidelity datasets. The playground shows their predictions next to the finite-element results, so the differences between architectures can be explored directly, and it runs the surrogates live in your browser on a case you set up yourself. You can find the project resources below:

Playground

Compare the surrogates with FEM on held-out test cases, or run them live in your browser on a case you set up.

Datasets

5,000 high-fidelity phase-field simulations in two datasets, openly available in the JHU Data Archive.

Paper

Fracture-Bench: a controlled benchmark suite for machine-learning surrogates in phase-field fracture mechanics.

Trained models

Weights of FNO, DeepONet and Latent DeepONet for both datasets, ready to load in JAX.

Benchmark poster

The datasets, the surrogates and the main results of the benchmark on one page.

Team

Department of Civil and Systems Engineering and Data Science and AI Institute, Johns Hopkins University, Baltimore, MD 21218, USA