[Paper Review] Parametrization and generation of geological models with generative adversarial networks

Shing Chan, Ahmed H. Elsheikh|arXiv (Cornell University)|Aug 5, 2017

Generative Adversarial Networks and Image Synthesis16 references62 citations

TL;DR

The paper uses Wasserstein GANs to parametrize geological models, producing samples that preserve multipoint statistics and flow behavior, often outperforming PCA in complex channelized patterns.

ABSTRACT

One of the main challenges in the parametrization of geological models is the ability to capture complex geological structures often observed in the subsurface. In recent years, generative adversarial networks (GAN) were proposed as an efficient method for the generation and parametrization of complex data, showing state-of-the-art performances in challenging computer vision tasks such as reproducing natural images (handwritten digits, human faces, etc.). In this work, we study the application of Wasserstein GAN for the parametrization of geological models. The effectiveness of the method is assessed for uncertainty propagation tasks using several test cases involving different permeability patterns and subsurface flow problems. Results show that GANs are able to generate samples that preserve the multipoint statistical features of the geological models both visually and quantitatively. The generated samples reproduce both the geological structures and the flow statistics of the reference geology.

Motivation & Objective

• Address the challenge of parametrizing complex subsurface geological models.
• Assess the effectiveness of GANs for parametric representation and uncertainty propagation.
• Compare GAN-based parametrization with PCA in reproducing higher-order statistics.
• Demonstrate GAN applicability to two subsurface flow problems and evaluate flow statistics.

Proposed method

• Use a Wasserstein GAN with a critic to minimize the Wasserstein distance between data and generated samples.
• Train GAN on permeability realizations cropped from 250x250 conceptual images to 50x50 realizations for two patterns (semi-straight and meandering).
• Experiment with input noise vector sizes of 20 and 40, using standard normal priors and tanh output activation.
• Preprocess binary permeability data to fit GAN training, and compare with PCA retaining 75% of variance (37 components semi-straight, 104 components meandering).
• Evaluate generated realizations through uncertainty propagation in two flow problems and compare flow statistics to reference data.

Experimental results

Research questions

• RQ1 Can GANs generate geological realizations that preserve multipoint statistics of reference models?
• RQ2 How does GAN-based parametrization compare to PCA in capturing complex channelized patterns?
• RQ3 Do GAN-generated realizations reproduce the flow statistics of the reference geology?
• RQ4 Are GAN-based samples effective for uncertainty propagation in subsurface flow problems?

Key findings

• GAN realizations visually capture channelized structures better than PCA, especially for meandering patterns.
• GANs reproduce the binary-like distribution of center-permeability values similar to data, while PCA yields near-normal distributions.
• In uncertainty propagation, GANs yield mean and variance close to true maps and better higher-order moments (skewness, kurtosis) than PCA.
• Densities of water breakthrough times estimated from GAN samples align with reference densities and outperform PCA.
• GANs achieve strong performance with 20 latent dimensions, outperforming PCA that uses 37 or 104 components for similar variance capture.