Graph Neural Networks for Reservoir Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current well performance models face challenges in efficiently simulating large hydrocarbon reservoirs due to increased runtime with multiple wells, and existing machine-learning approaches often trade accuracy for speed.
Innovation Solution
The use of a graph neural network (GNN) generated from a reservoir graph network, where grid cells and coarsened grid blocks are associated with nodes, and dynamic responses from previous simulations determine graph edges, allowing for faster and more accurate reservoir simulations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the number of wells in a reservoir simulation is increased, then the accuracy and completeness of reservoir modeling is improved, but the runtime necessary to perform the reservoir simulation increases
Solution Approach 1:
The patent creates a simplified copy of the reservoir system using graph neural networks that replicate the behavior of the full physics-based reservoir simulation. The GNN model learns from training data generated by traditional reservoir simulations and can predict well performance without requiring the computationally intensive full simulation process, thus providing accurate predictions with significantly reduced runtime
Solution Approach 2:
The patent replaces the mechanical physics-based numerical simulation system with a machine learning-based graph neural network system. Instead of solving complex partial differential equations that govern fluid flow in reservoirs, the system uses a GNN that has been trained to approximate these physical processes, substituting the traditional computational mechanics approach with an intelligent system that achieves similar accuracy much faster
2Measurement precision
If traditional physics-based reservoir modeling is used, then simulation accuracy is maintained, but computational speed decreases
Solution Approach 1:
The patent performs preliminary actions by training the graph neural network model in advance using data from traditional physics-based reservoir simulations. During the training phase, the GNN learns the complex relationships between reservoir properties, well configurations, and production outcomes. Once trained, the model can make predictions without requiring the full physics simulation process, achieving both accuracy and speed
Solution Approach 2:
The patent changes the fundamental parameters of the simulation approach by transitioning from continuous physics-based numerical methods to a discrete graph-based machine learning representation. The reservoir is represented as a graph where nodes represent grid cells and edges represent connectivity, with properties encoded as node and edge features. This parameter transformation enables the use of efficient GNN algorithms that can process reservoir data much faster than traditional solvers while maintaining predictive accuracy
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A method may include obtaining dynamic response data regarding a geological region of interest. The dynamic response data may include various transmissibility values. The method may further include determining a reservoir graph network based on the dynamic response data and a reservoir grid model. The reservoir graph network may include various grid cells, various wells, and various graph edges. The method may further include generating a graph neural network based on the reservoir graph network. The method may further include updating the graph neural network using a machine-learning algorithm to produce an updated graph neural network for the geological region of interest. The method may further include simulating a well within the geological region of interest using the updated graph neural network.