Latent Diffusion Model for Conditional Reservoir Facies Generation

TL;DR

A novel Latent Diffusion Model is proposed, which is specifically designed for conditional generation of reservoir facies, and significantly outperforms a GAN-based alternative.

Abstract

Creating accurate and geologically realistic reservoir facies based on limited measurements is crucial for field development and reservoir management, especially in the oil and gas sector. Traditional two-point geostatistics, while foundational, often struggle to capture complex geological patterns. Multi-point statistics offers more flexibility, but comes with its own challenges related to pattern configurations and storage limits. With the rise of Generative Adversarial Networks (GANs) and their success in various fields, there has been a shift towards using them for facies generation. However, recent advances in the computer vision domain have shown the superiority of diffusion models over GANs. Motivated by this, a novel Latent Diffusion Model is proposed, which is specifically designed for conditional generation of reservoir facies. The proposed model produces high-fidelity facies realizations that rigorously preserve conditioning data. It significantly outperforms a GAN-based alternative. Our implementation on GitHub: \url{https://github.com/ML4ITS/Latent-Diffusion-Model-for-Conditional-Reservoir-Facies-Generation}.

Figures (13)

• Figure 1: Illustration of our conditional reservoir generation problem in which the generative model stochastically samples a realistic reservoir (right) given the limited measurements (left). The regions with no information are denoted in grey.
• Figure 2: Illustration of the U-Net architecture cai2022novel, where Conv denotes a convolutional layer. U-Net is a convolutional neural network architecture, featuring an encoder (first half of U-Net) and decoder (second half) structure with skip connections that allow for the transfer of spatial information across layers, which in turn enables precise localization and high-resolution output.
• Figure 3: Illustration of the principle of a diffusion process. The diffusion modeling mainly consists of 1) forward process (noising) and 2) reverse process (denoising). The noising process begins with a data sample and incrementally adds Gaussian noise over multiple time steps to convert it into a Gaussian noise sample; conversely, the denoising process iteratively refines this Gaussian noise sample back into a data-like sample, guided by a neural network trained specifically for this denoising task.
• Figure 4: Overview of LDM. The encoder $\mathcal{E}$ and decoder $\mathcal{D}$ enable data compression, enabling the forward and reverse processes to operate in a reduced-dimensional space. This eases the task of learning prior and posterior distributions and improves computational efficiency. In addition, conditional data can be fed into the reverse process, enabling conditional generation.
• Figure 5: Overview of our proposed method. Our method can be regarded as an adapted version of LDM to effectively handle the categorical input and allow maximal preservation of conditional facies data in generated facies while maintaining the high fidelity of generated facies.