Xupeng He
Petroleum Engineer
Saudi Aramco
Participates in
TECHNICAL PROGRAMME | Energy Technologies
The Energy Transition: The Role of Digitalisation, AI, and Cybersecurity
Forum 23 | Digital Poster Plaza 4
30
April
10:00
12:00
UTC+3
Objective/Scope:
Calibrating simulation models and estimating uncertain parameters heavily relies on history matching. However, the presence of subsurface uncertainties often makes this process highly intricate. To tackle this complexity, we present a new methodology that combines Bayesian inversion with the Coarse-grid Network model, offering a more efficient and streamlined path for history matching.
Methods, Procedures, Process:
The process follows a structured five-step framework. Initially, key variables such as injection rate, porosity, capillary pressure, oil viscosity, density, and relative permeability are selected through Global Sensitivity Analysis (GSA). Their initial ranges and statistical distributions are established based on prior information. In the second stage, Latin Hypercube Sampling (LHS) is applied to generate both training and testing datasets, which are then simulated using a high-resolution numerical model. A Coarse-grid Network (CgNet) model is subsequently developed to learn the input-output relationships of these parameters, while Bayesian optimization is employed to fine-tune the model's hyperparameters automatically. In the fourth step, real-world field data are incorporated into the Bayesian inversion framework, where the Markov Chain Monte Carlo (MCMC) algorithm is used to compute posterior distributions of the uncertain parameters. Lastly, the resulting parameter estimates are verified by comparing the simulated pressure outcomes with the actual observations. Should significant mismatches occur, the performance of the CgNet surrogate model is reassessed and enhanced accordingly.
Results, Observations, Conclusions:
The workflow is applied to the COSTA model, which accounts for a range of physical phenomena, including capillary, viscous, and gravitational forces. This model features a high-resolution grid of 100 million cells and includes 240 injection and production wells, all operating under realistic control scenarios. The results demonstrate that the integration of Bayesian inversion with the CgNet model successfully identifies uncertain parameters, significantly narrowing the range of subsurface uncertainties. The difference between the actual and estimated pressure responses is minimal, highlighting the accuracy of the method. Additionally, when compared to other surrogate-based Bayesian inversion techniques—such as kriging, support vector machines, and polynomial chaos expansion—the CgNet-based method shows improved performance. This enhancement is attributed to CgNet's strength in handling temporal data efficiently, making it particularly well-suited for the dynamic nature of reservoir simulation and history matching.
Calibrating simulation models and estimating uncertain parameters heavily relies on history matching. However, the presence of subsurface uncertainties often makes this process highly intricate. To tackle this complexity, we present a new methodology that combines Bayesian inversion with the Coarse-grid Network model, offering a more efficient and streamlined path for history matching.
Methods, Procedures, Process:
The process follows a structured five-step framework. Initially, key variables such as injection rate, porosity, capillary pressure, oil viscosity, density, and relative permeability are selected through Global Sensitivity Analysis (GSA). Their initial ranges and statistical distributions are established based on prior information. In the second stage, Latin Hypercube Sampling (LHS) is applied to generate both training and testing datasets, which are then simulated using a high-resolution numerical model. A Coarse-grid Network (CgNet) model is subsequently developed to learn the input-output relationships of these parameters, while Bayesian optimization is employed to fine-tune the model's hyperparameters automatically. In the fourth step, real-world field data are incorporated into the Bayesian inversion framework, where the Markov Chain Monte Carlo (MCMC) algorithm is used to compute posterior distributions of the uncertain parameters. Lastly, the resulting parameter estimates are verified by comparing the simulated pressure outcomes with the actual observations. Should significant mismatches occur, the performance of the CgNet surrogate model is reassessed and enhanced accordingly.
Results, Observations, Conclusions:
The workflow is applied to the COSTA model, which accounts for a range of physical phenomena, including capillary, viscous, and gravitational forces. This model features a high-resolution grid of 100 million cells and includes 240 injection and production wells, all operating under realistic control scenarios. The results demonstrate that the integration of Bayesian inversion with the CgNet model successfully identifies uncertain parameters, significantly narrowing the range of subsurface uncertainties. The difference between the actual and estimated pressure responses is minimal, highlighting the accuracy of the method. Additionally, when compared to other surrogate-based Bayesian inversion techniques—such as kriging, support vector machines, and polynomial chaos expansion—the CgNet-based method shows improved performance. This enhancement is attributed to CgNet's strength in handling temporal data efficiently, making it particularly well-suited for the dynamic nature of reservoir simulation and history matching.
