This repository contains the code used to generate the results of the research article
David Melching, Erik Schultheis and Eric Breitbarth. (2024)
Generating artificial displacement data of cracked specimen using physics-guided adversarial networks
Machine Learning: Science and Technology.
DOI: 10.1088/2632-2153/ad15b2
The published article is available here. The preprint is available on arXiv.
Digital image correlation (DIC) has become a valuable tool to monitor and evaluate mechanical experiments of cracked specimen, but the automatic detection of cracks is often difficult due to inherent noise and artefacts. Machine learning models have been extremely successful in detecting crack paths and crack tips using DIC-measured, interpolated full-field displacements as input to a convolution-based segmentation model. Still, big data is needed to train such models. However, scientific data is often scarce as experiments are expensive and time-consuming. In this work, we present a method to directly generate large amounts of artificial displacement data of cracked specimen resembling real interpolated DIC displacements. The approach is based on generative adversarial networks (GANs). During training, the discriminator receives physical domain knowledge in the form of the derived von Mises equivalent strain. We show that this physics-guided approach leads to improved results in terms of visual quality of samples, sliced Wasserstein distance, and geometry score when compared to a classical unguided GAN approach.
- Python 3.11.5
- For needed python packages see requirements.txt
This repository provides different GAN architectures to generate DIC displacement data aswell as metrics for evaluation.
This repository was mainly used to evaluate two different GAN types:
- A GAN with a generator which outputs displacement data and a discriminator taking displacement data as input (Classical GAN)
- A GAN with a generator which outputs displacement data and a discriminator taking displacement data and pixel-wise calculated von-Mises equivalent strain data as input (Physics-guided GAN).
The GAN API is found in the gan module.
Different discriminators and generator are defined in the files discriminators.py and generators.py, respectively.
layers.py implements some costum layers.
gans.py implements functions and classes for creating, training, saving and loading GAN models.
train_gan.py and test_gan.py show how to create, train, save, load
and evaluate a GAN.
We use the GAN class in gan.gans to create a GAN.
For this, we need a discriminator and a generator (e.g. from gan.discriminators and gan.generators).
Further, a loss object from the class gan.gans.GANLoss defines the used loss and regularization
for the training process. Additional arguments are the used optimizers for both discriminator and generator.
The device (CPU or GPU) and whether to use Automated Mixed Precision (AMP)
can be specified with the corresponding arguments.
The argument strains is used to specify which strains are calculated
and used as additional input for the discriminator while training.
When training from scratch use the gan.gans.init_weights function to initialize
the weights of a GAN. We train a GAN with its train function. train_gan.py is a script for creating, training and saving a GAN which uses configs.gan for configuration.
The training data used in the publication can be found on Zenodo.
While training, a GAN is automatically saved after a certain number of epochs
(checkpoint_rate argument of the train function) and at the end of training.
We can manually save a GAN with its save_gan function.
A GAN is saved in a GAN state_dict. This state_dict combines the state_dicts
of the generator, discriminator, loss and optimizers.
The name of the folder in which a GAN is saved,
characterizes a certain GAN.
The folder name contains hyperparameters of the training process and the
used architecture.
We load a trained GAN with the gan.gans.load_gan_model_by_folder function.
This function creates all necessary components to create a GAN object. test_gan.py is a script for loading and evaluating a trained GAN. All GANs trained with the provided GAN API or the train_gan.py script are loaded with the gan.gans.load_gan_model_by_folder function`.
This repository implements the Structural-Similarity-Index-Measure (SSIM), Multiscale-Structural-Similarity-Index-Measure (MS-SSIM), Sliced Wasserstein Distance (SWD) and the Geometry Score (GS) as metrics to evaluate the generated fake data of a GANs generator. We implement a costum version of the SWD in PyTorch which is able to utilize the GPU (see metrics.swd_torch). In this project, mainly, the SWD and GS were used to evaluate the trained GANs. Both metrics reveal mode collapse and was used to quantify quality and variation of fake data generated with a trained GAN.
test_gan.py shows examples of plotting fake and real data as well as plotting metrics. For plotting data we use functions from plot.data. For plotting metrics we use functions from plot.metrics. plot.data.plot_batch is used to plot and save a batch of data. Plotting single samples, real versus fake comparisons and a series of data as an animation can be done with plot.data.plot_sample, plot.data.plot_real_vs_fake and plot.data.plot_batch_animation respectively.
For plotting SWDs and MRLTs (from the geometry score) for multiple trained models, we recommend to first save the results of multiple models with the metrics.swd.save_swds and metrics.geometry_score.save_geom_scores functions and then plot the results with plot.metrics.plot_swds and plot.metrics.plot_geom_scores.
For plotting statistics like the loss values we use the plot.metrics.plot_statistics function.