Geological realism in Fluvial facies modelling with GAN under variable depositional conditions
by Chao Sun, Vasily Demyanov, Daniel Arnold
Computational Geosciences, 10.1007/s10596-023-10190-w
This study investigates generative adversarial networks (GANs)’ capacity to model multi-facies distributions of meandering systems. Earlier works showed that GANs outperform geostatistical methods in reproducing complex geometry, like the shapes of fluvial channels. However, the reproduction of geological complexity and geological realism remains an issue when modelling fluvial depositional systems. Meandering systems deposit multiple facies and change facies shape following the migration of rivers. Sand accretes at the inner bank of channels, forming the point bar and erodes the plain at the outer bank to create sediments. Channel fills with mud or sand at the bottom after abandonment due to avulsions or meander cutoffs. Those sedimentary processes yield complex geological patterns.
This paper proposes further developing a GAN model, Fluvial GAN, to learn complex multi-facies fluvial patterns across depositional variability. We create a set of meandering facies models by a process-based model, FLUMY (TM), for training a GAN and assessing how well it can learn fluvial facies distributions representing sedimentary processes.
Our fluvial GAN has three distinct enhancements:
- a One-Hot Encoder for better handling of multi-facies distribution,
- a Hybrid-discriminator for better learning geological patterns, and
- an improved loss function to prevent mode collapse.
We compare Fluvial GAN performance with two more standard configurations using qualitative and quantitative geological features assessments. Fluvial GAN vastly reduces the occurrence of a typical unrealistic feature, channels forming isolated loops, which we called ‘closed channel’ in this study. We analyse the diversity of Fluvial GAN generations via a dimensionality reduction algorithm, UMAP, that plots the training dataset and Fluvial GAN generations together in a 2D space. Fluvial GAN provides good coverage of the uncertainty space represented by the training dataset.