[Paper Review] Latent Diffusion Model for Conditional Reservoir Facies Generation

Dae-Soo Lee, Ovanger, Oscar+4 more|arXiv (Cornell University)|Nov 3, 2023

Reservoir Engineering and Simulation Methods4 citations

TL;DR

This paper proposes a novel Latent Diffusion Model (LDM) for conditional reservoir facies generation, leveraging a denoising U-Net in latent space to produce high-fidelity, geologically realistic facies realizations that strictly preserve conditioning data. The method significantly outperforms a GAN-based baseline in fidelity, sample diversity, and conditional data preservation, marking the first application of diffusion models to this problem in reservoir modeling.

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}.

Motivation & Objective

• To address the limitations of traditional geostatistical and GAN-based methods in generating realistic, conditionally constrained reservoir facies realizations.
• To explore the potential of diffusion models—particularly latent diffusion models—for conditional facies generation in reservoir modeling.
• To develop a method that preserves conditioning data with high fidelity while generating diverse, realistic facies patterns.
• To demonstrate the superiority of LDM over GAN-based approaches in terms of generative quality and preservation accuracy.
• To establish a new benchmark for conditional reservoir facies modeling using state-of-the-art generative AI techniques.

Proposed method

• The model employs a latent diffusion framework where the diffusion process occurs in a compressed latent space, improving computational efficiency while maintaining high-fidelity generation.
• A denoising U-Net is trained to reverse the diffusion process, conditioned on both the latent noise and a latent representation of conditioning data.
• The conditioning data are encoded into a latent vector $\mathbf{z}_c$ via a trainable encoder, which is then integrated into the denoising U-Net through multi-level feature concatenation.
• A novel preservation loss $L_{\text{preserv}}$ is introduced to explicitly penalize deviations from the conditioning data during training, enhancing fidelity to observed measurements.
• The model is trained on 5,000 synthetic 2D facies realizations, with the generative process framed as a conditional generation problem where the model learns to generate realistic samples given limited data.
• Ablation studies validate the necessity of key components: the trainable discriminator $\mathcal{D}$, the preservation loss $L_{\text{preserv}}$, and the multi-level mapping of $\mathbf{z}_c$.

Experimental results

Research questions

• RQ1 Can a latent diffusion model effectively generate high-fidelity, geologically realistic reservoir facies realizations conditioned on sparse measurements?
• RQ2 How does the proposed LDM compare to a GAN-based baseline in terms of conditional data preservation, sample diversity, and generative fidelity?
• RQ3 What role do the preservation loss $L_{\text{preserv}}$, trainable discriminator $\mathcal{D}$, and multi-level mapping of $\mathbf{z}_c$ play in enhancing conditional preservation?
• RQ4 To what extent does the LDM overcome the limitations of GANs, such as mode collapse and training instability, in reservoir facies modeling?
• RQ5 Can the LDM be extended to 3D reservoir modeling, and what challenges arise from the increased dimensionality?

Key findings

• The proposed LDM achieved a preservation error rate of 0.0003 in the base case, indicating near-perfect preservation of conditioning data, significantly outperforming the GAN baseline.
• The LDM achieved a reconstruction loss $L_{\text{recons}}$ of 0.0028 and a total loss $L_{\text{LDM}}$ of 0.0236, demonstrating high-fidelity generation.
• Removing the preservation loss $L_{\text{preserv}}$ increased the preservation error rate to 0.0094, confirming its critical role in maintaining conditioning fidelity.
• The absence of a trainable discriminator $\mathcal{D}$ led to a $L_{\text{LDM}}$ increase to 0.0438 and a preservation error rate of 0.011, highlighting its importance in learning realistic distributions.
• Replacing multi-level mapping with simple concatenation increased the preservation error rate to 0.0017 and raised $L_{\text{LDM}}$ to 0.0262, proving the effectiveness of hierarchical feature integration.
• The ablation study confirmed that all proposed components—$L_{\text{preserv}}$, trainable $\mathcal{D}$, and multi-level mapping—significantly enhance the model’s ability to preserve conditioning data.