4.5 Article

Progressive growing generative adversarial networks using conditioning ratio for facies modeling in complex aquifers

Journal

HYDROGEOLOGY JOURNAL
Volume -, Issue -, Pages -

Publisher

SPRINGER
DOI: 10.1007/s10040-023-02687-6

Keywords

Generative adversarial networks; Facies modeling; Machine learning; Geological modeling; Geostatistics

Ask authors/readers for more resources

Groundwater-flow and contaminant-transport modeling can benefit from the use of machine learning techniques such as generative adversarial networks (GANs). This study focuses on a progressive growing GAN (PGGAN) to generate geologically realistic images of channel aquifers based on field observations. The conditioning behavior and its influence on the network architecture were measured using a novel metric called the conditioning ratio. The results revealed different conditioning behaviors based on the number of conditioning arrays injected into the generator.
Groundwater-flow and contaminant-transport modeling rely on methods of converting a set of field observations into geologic models that represent the subsurface structure. These geologic models also must replicate important geologic features such as connectivity. Recently, researchers have begun to use machine learning methods such as generative adversarial networks (GANs). This study focuses on a progressive growing GAN (PGGAN) to condition on measured data. Given a latent variable and an array that provides field observations, the generators of the conditioned PGGAN are tasked to produce geologically realistic images of channel aquifers that match field observations. Although largely successful, the conditioning behavior of these networks still has some issues, and how the model performs the conditioning task across its layers is not yet fully understood. To better understand this conditioning mechanism, the behavior of these networks was measured using the conditioning ratio, which is a novel metric that determines the magnitude of the influence of the conditioning data. The conditioning ratio was measured across multiple layers within the generator during training, as well as with various modifications to the network architecture. The results revealed two distinct conditioning behaviors that are based on the number of conditioning arrays injected into the generator. Results also showed that decreasing the starting resolution for the generator can slow down the learning process. Overall, the numerical experiments prove the value of measuring the conditioning ratio of layers within the generator. These approaches can be used as diagnostic tools to assist in the design of future PGGAN architectures.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.5
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available