Browsing by Author "Abdellatif, Alhasan"

Now showing 1 - 1 of 1

Examining the generalization performance of Generative Adversarial Networks in geology
(Heriot-Watt University, 2023-10) Abdellatif, Alhasan; Elsheikh, Ahmed H.
This thesis investigates the generalization capabilities of Generative Adversarial Networks (GANs) in geology. The main motivation is to utilize the expressive power of deep generative models to address the challenges in synthesizing realistic and diverse geological realizations particularly when presented with limited training set. The thesis delves into the architecture and training methodologies of GANs tailored to geology-specific characteristics, such as geological feature representation within the given samples, spatial dependencies (e.g., non-stationarity), and large-scale generation. The thesis first describes how GANs can generate geological channelized patterns with global proportions beyond a training set that lacks representative samples. The work modifies the standard conditioning techniques to accommodate for the missing patterns in the training set. The second work extended the first method to work in a spatial setting, where it can be trained to generate non-stationary patterns. It uses a spatial conditioning method and proves to generate geologically-consistent samples that respect target probability maps. The final work introduces a novel approach for generating texture patterns of arbitrary large sizes, including geological samples, given a single small-resolution image (e.g., 256 × 256). It uses a patch-by-patch generating technique that efficiently utilizes the GPU resources and is able to generate coherent large-resolution texture images.