Reservoir History Matching of the Norne field with generative exotic priors and a coupled Mixture of Experts -- Physics Informed Neural Operator Forward Model

Overview

• This paper presents a novel approach for reservoir history matching using generative exotic priors and a coupled Mixture of Experts - Physics Informed Neural Operator Forward Model.
• The method is applied to the Norne field, a North Sea oil and gas reservoir, to improve characterization and prediction of its subsurface properties.
• The key innovations include the use of generative adversarial networks to create realistic prior distributions, and a physics-informed neural network architecture to model the complex reservoir dynamics.

Plain English Explanation

Reservoir history matching is the process of adjusting the computer model of an oil or gas reservoir to match the actual observed behavior of the reservoir over time. This is important for accurately predicting future production and optimizing extraction strategies.

The researchers in this paper developed a new approach to improve reservoir history matching. They used a type of machine learning called generative adversarial networks (GANs) to create realistic "prior" distributions of possible reservoir properties, like the porosity and permeability of the rocks. These prior distributions serve as the starting point for the history matching process.

The researchers also used a special type of neural network architecture called a "Physics Informed Neural Operator" to model the complex physical processes happening in the reservoir. This allowed the model to better capture the underlying physics governing the reservoir's behavior, rather than just relying on historical data.

By combining the realistic prior distributions from the GANs and the physics-informed neural network model, the researchers were able to more accurately match the historical behavior of the Norne oil and gas field in the North Sea. This can lead to better predictions of future production and more effective extraction strategies.

Technical Explanation

The paper introduces a novel framework for reservoir history matching that leverages generative adversarial networks (GANs) to create realistic prior distributions of reservoir properties, and a Physics Informed Neural Operator (PINO) to model the forward dynamics of the reservoir.

The GAN-based prior is used to generate a diverse set of plausible reservoir realizations, which are then history matched against observed production data using a Mixture of Experts (MoE) architecture. The MoE model consists of multiple neural networks, each specializing in different regions or flow regimes of the reservoir, coupled with a gating network that learns to adaptively combine their predictions.

The forward model within the MoE framework is a PINO, which incorporates the underlying physics of the reservoir dynamics into the neural network architecture. This PINO forward model is trained on high-fidelity simulations to learn an efficient approximation of the reservoir's behavior.

The combined GAN-based prior and PINO-based forward model is applied to the Norne field, a complex North Sea oil and gas reservoir, demonstrating improved history matching performance compared to traditional approaches.

Critical Analysis

The paper presents a comprehensive and innovative approach to reservoir history matching, addressing key challenges in this domain. The use of generative adversarial networks to create realistic prior distributions is a particularly compelling aspect, as it allows for the exploration of a broader range of plausible reservoir scenarios.

However, the authors acknowledge that the GAN-based prior may not fully capture the underlying geological complexity and heterogeneity of real-world reservoirs. Additionally, the training of the PINO forward model requires access to high-fidelity reservoir simulations, which may not always be readily available or computationally feasible for large-scale, high-resolution models.

Further research could explore ways to relax these requirements, such as developing techniques to learn the physics-based forward model directly from field data or incorporating additional geological and geophysical information into the prior distribution. Additionally, the scalability and computational efficiency of the proposed framework for large-scale, complex reservoirs could be further investigated.

Conclusion

This paper presents a novel approach for reservoir history matching that combines generative adversarial networks for creating realistic prior distributions and a Physics Informed Neural Operator forward model within a Mixture of Experts framework. The method is applied to the Norne field and demonstrates improved history matching performance compared to traditional techniques.

The use of GAN-based priors and physics-informed neural networks represents a significant advancement in the field of reservoir characterization and could lead to more accurate predictions of future production, ultimately supporting more effective extraction strategies and better decision-making in the oil and gas industry.