Chao Sun successfully passed his viva with a thesis on GAN fluvial facies distribution modelling

Congratulations to Chao Sun with his successful PhD viva as the result of 4 years of hard work as a NERC CDT in Oil and Gas PhD student supervised by prof. V. Demyanov and  Dr. D. Arnold.
The examiners – Prof Guillaume Rongier (TU Delft) and Prof. Ahmed Elsheikh (HWU) have assessed the work in a great detail, at high level of rigour
Use of Generative Learning to Improve Realism in Fluvial Facies Modelling,

which has been already validated through 2 published journal papers with Computational Geosciences and Data in Brief as well as 1 more in review with Computer and Geosciences, along with many conference papers with EAGE and IAMG.

This thesis investigates using generative adversarial networks (GANs) to efficiently build geologically plausible 3D facies models of fluvial systems by learning realistic variations of realistic facies patterns and their uncertainty from a process-based model. Fluvial systems, e.g. meandering rivers, often create complicated facies distribution composed of multiple facies with varied shapes and transitions due to the complex sedimentary processes. Conventional simulation tools, such as process-based models and geostatistical approaches, use a stochastic process to simulate fluvial facies models based on physic based or rule-based processes, parametric geometries or spatial correlation models.
Deep generative models, e.g. GANs, showed powerful learning capability that allows using limited latent parameters to sample random realisations, naturally a geological parameterisation. GANs have successfully reproduced realisations from object-based models. This triggers the interest in exploiting the learning capacity for data complexity. As GANs can learn geological patterns from object-based models, how about process-based models and even nature geology?

This thesis exploited GAN’s ability to learn geologically realistic facies distributions  from a process-based simulator, FLUMY, and demonstrate the value of deep generative modelling in real-world subsurface challenges. This work tackled several identified problems in GAN learning 2D and 3D meandering fluvial patterns by proposing a set of unique model structures, learning frameworks and training strategies. The ultimate product of this PhD project is a GAN-based 3D facies modelling tool for low net-to-gross meandering fluvial systems called FluvialGAN3D simulator.
The Fluvial-GAN3D simulator consists of two pre-trained generators and a reconstruction program, achieved by solving the problems below in the thesis:

  1. creating a benchmark meandering fluvial multi-facie dataset GAN River I available for reproduction for UQ tasks, that represents the features and the variability of the process-based simulations;
  2. comparison of different GAN setups;
  3. efficient GAN training on 2D patterns to reconstruct 3D facies models;
  4. geological consistent 3D reconstruction of the deposited succession of arbitrary thickness;
  5. investigating different GAN extensions, including soft conditioning to well and seismic data.
Chao has shared the findings of his PhD at a EAGE and IAMG conferences, NERC CDT annual conferences and of course with the GeoDataScience Group Industry Knowledge Share Days that provided insights into his work.

Congratulations to Chao – it’s been a great privilege to advise such a bright and hard working student,  who has demonstrated the most impressive steep learning curve!
And best wishes for his further career.