Iterative Proxy Model Construction for Reservoir Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current reservoir simulation methods, such as Monte Carlo simulations and proxy models, face challenges in efficiently handling large dimensional parameter spaces with non-linearities, leading to high computational costs and inaccurate predictions due to limitations in capturing non-linear effects and probability distributions.
Innovation Solution
The method involves constructing and refining proxy models using iterative sampling techniques, specifically selecting new sampling points based on response surface properties like value, gradient, curvature, and distance, to capture non-linearities and improve model accuracy, employing thin-plate spline regression models to approximate reservoir simulation outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Monte Carlo simulation is used for uncertainty quantification, then statistical accuracy of results is improved, but computational cost increases prohibitively
Solution Approach 1:
The patent creates a proxy model (a simplified copy) that replicates the behavior of the complex reservoir simulation model. This proxy model can be evaluated rapidly without requiring full computational simulations, thus providing statistically accurate uncertainty quantification at a fraction of the computational cost of Monte Carlo methods.
Solution Approach 2:
The patent introduces a proxy model as an intermediary between the input parameters and the full reservoir simulation. This intermediary captures the essential input-output relationships and allows for efficient uncertainty analysis without repeatedly running computationally expensive simulations.
2Productivity
If proxy models are used to improve computational efficiency, then evaluation speed increases, but training requirements grow exponentially with parameter space dimension
Solution Approach 1:
The patent performs preliminary analysis to identify and focus training efforts on the most influential parameters and regions of the parameter space. By using adaptive sampling techniques, the method prepares training data in advance in a way that maximizes proxy model accuracy while minimizing the number of training points required, thus avoiding exponential growth in training complexity.
Solution Approach 2:
The patent applies adaptive sampling that concentrates training points in regions of the parameter space that are most important for model accuracy. Rather than uniformly distributing training points, the method allocates more computational resources to critical regions, thereby reducing overall training requirements while maintaining high evaluation speed.
3Ease of manufacture
If experimental design methods are used for proxy model construction, then model generation is simplified, but accuracy deteriorates when parameter distributions are arbitrary and non-linear effects are strong
Solution Approach 1:
The patent employs adaptive sampling techniques that dynamically adjust the placement of training points based on the evolving understanding of the parameter space. Rather than using fixed experimental design grids, the method adaptively concentrates samples in regions where the proxy model needs improvement, particularly where strong non-linear effects occur, thereby maintaining prediction accuracy without sacrificing simplicity.
Solution Approach 2:
The patent implements an iterative process where the proxy model is continuously refined based on feedback from additional simulations. The method identifies regions where the current model predictions are inaccurate and automatically adds training points in those regions, improving prediction accuracy while maintaining ease of model generation through automated adaptive sampling.
Data Source
AI summary
A method, system and computer program product is disclosed for utilizing proxy models to evaluate a subterranean reservoir. The method includes constructing a proxy model from a set of sampling points to approximate simulation outputs of a reservoir model. The set of sampling points is updated by adding at least one new sampling point that is selected from a location associated with surface non-linearities such as gradients, curvature, and bending energy. Response surface values at new sampling points and distances to existing sampling points can also be used to evaluate new sampling points. Proxy models are refined with the updated set of sampling points until the proxy model satisfies a predetermined stopping criterion, such as when a predetermined number of iterations are reached or when changes to the response surface are below a predetermined threshold.