Use of Generative Adversarial Networks to incorporate the Training Image Uncertainty in Multiple-Point Statistics Simulation

Objective • Results • Usage •

🎯 Objectives

Multiple-Point Geostatistical (MPS) methods have been successfully applied to build numerical models with curvilinear features using several sources of information. Even though traditional algorithms reproduce the spatial pattern of the variogram models, they fail describing curvilinear features - which came from a conceptual model of the underlying geology provided by the expert geologist. However, there is hope - MPS new methods reproduces these patterns we wish to replicate.

In this work, we chose the SNESIM algorithm (Strebelle, 2002) for three reasons: (a) because it is a method widely used by the community; (b) its parameters are intuitive and interpretable; (c) SNESIM algorithm is freely available (Remy and Boucher, 2009).

The training image is uncertain as the actual spatial pattern is unknown. This uncertainty is even more pronounced at the exploration stage when little information is available (Pyrcz and Deutsch, 2014). Considering the uncertainty of the input parameters improves the assessment of the space of uncertainty, Pyrcz and Deutsch (2014) recommend using a scenario-based approach to incorporate the lack of confidence of the parameters in the simulations.

The idea is to merge the generative model adeptness to learn spatial patterns with the benefits of the SNESIM algorithm to use many types of information for conditioning. The outcome is a hybrid workflow with two main steps: (a) creating a dataset of training images using the generative model; (b) building geostatistical models using the synthetic TI and the existing conditional data.

Results and discussion

👷 Usage

Pre-requisites to run the script included in the requirements.txt file .

git clone https://github.com/algocompretto/gan-uncertainty.git

# Activates the environment and installs prerequisites
cd gan-uncertainty/ && python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Running SNESIM simulations

In the project folder, navigate to the `SNESIM` folder, and then execute the script with:

python3 snesim.py --arguments

Argument name	Description
--samples_path	The samples path. The file should contain data in the following format: x,y,z,facies .
--ti_path	The training image path in GSLIB format.
--par_path	Path to the parameter file with all information related to the simulation process itself.
--exe_path	The snesim.exe file path.
--output_path	Path to the output file.
--realizations	Number of realizations to be done.
--max_cond	The maximum amount of points to use in the conditioning process.
--min_cond	The minimum amount of points to use in the conditioning process.
--seed	The initial seed for the simulation.
--plot	A boolean value for whether you want to plot/save the results or not.

Running the proposed workflow

Training

If you wish to train a new model on unseen data, you can follow the next steps:

cd generative_model/
python3 gan.py

The training will get all the information on hyperparameters from the parameters.yaml file

Argument name	Description
output_dir	The output directory for the augmented images.
training_image	The training image path in .png format.
checkpoint	The checkpoint folder which the models will be stored.
sample_images	The folder where the sampled examples from the network will be saved.
num_channels	Number of channels in the image
latent_dim	The latent dimension vector size representing the features.
learning_rate	The learning rate for the Adam optimizers
images_path	The output directory for the windowed images.
batch_size	The batch size for the training step.
num_workers	The number of workers to load the dataset.
num_epochs	The number of epochs for training step.
cuda	A boolean value for whether you want to use the CUDA device or not.
n_critic	The number of steps to train the Critic after n iterations of the Generator.