Petroleum System Model Calibration Using ML Surrogate Simulations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Basin and petroleum system modeling faces economic impracticality due to high costs associated with computationally intensive ensemble-based simulations, which are often restricted to a limited number of realizations, limiting the applicability of petroleum system modeling.
Innovation Solution
A method utilizing machine-learning models to predict simulation outcomes based on input parameters and realizations, allowing for the selection of candidate simulations and output parameters that minimize a target function, thereby reducing the need for extensive ensemble simulations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ensemble-based simulations with high number of realizations are performed to accurately describe uncertainty, then prediction accuracy is improved, but computational cost increases making the approach economically impractical
Solution Approach 1:
The patent creates a surrogate model (copy) of the complex petroleum system simulator using machine learning. This surrogate model approximates the behavior of the full ensemble simulations but requires significantly fewer computational resources, enabling accurate uncertainty quantification without the prohibitive cost of running thousands of full simulations
Solution Approach 2:
The patent transforms the input parameters from the original simulation space into a reduced-dimensional latent space using dimensionality reduction techniques. This parameter transformation allows the machine learning model to capture the essential variability with far fewer parameters, reducing computational cost while maintaining prediction accuracy
2Ease of manufacture
If the number of simulation realizations is restricted to a limited number, then computational cost is reduced, but the applicability and reliability of petroleum system modeling deteriorates
Solution Approach 1:
The patent implements an active learning framework where the surrogate model is iteratively improved by selecting new training points based on prediction uncertainty. This feedback loop allows the model to achieve high reliability with a limited number of expensive full simulations, as the system automatically identifies which additional simulations would provide the most value
Solution Approach 2:
The patent performs preliminary dimensionality reduction and feature extraction on the input parameters before running simulations. This preliminary processing organizes the parameter space in a way that maximizes the information gained from each simulation, allowing reliable predictions with fewer realizations
Data Source
AI summary
A method for simulating a subterranean volume includes receiving one or more input parameters and one or more simulation realizations representing the subterranean domain, modeling the one or more simulation realizations as a target function of the one or more input parameters, training a machine-learning model to predict values for the target function using the one or more input parameters and the one or more simulation realizations, predicting a value for the target function based on a first candidate simulation or a first candidate output parameter of a simulation, selecting the first candidate simulation, the first candidate output parameter, or both based on the predicted value of the target function, and simulating the subterranean volume using the first candidate output parameter, or both.