Machine-Learning Surrogate Models for Reservoir Phase Equilibria
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Compositional reservoir simulations face inefficiencies in thermodynamic calculations, particularly in stability analysis and phase-split calculations, which are often iterative and require extensive data processing, limiting accuracy and computational speed.
Innovation Solution
Training machine-learning-based surrogate models using reservoir-specific compositional databases generated from PVT experiments, allowing for non-iterative phase stability and flash calculations by interpolating fluid compositions across the reservoir and simulating physical processes like depletion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative thermodynamic calculations are used for stability analysis and phase-split calculations, then accuracy of phase behavior prediction is improved, but computational time and processing speed deteriorate
Solution Approach 1:
The patent pre-calculates and stores thermodynamic properties (phase stability, flash calculations) in lookup tables before reservoir simulation begins. These pre-computed results are organized by composition, pressure, and temperature, allowing the simulation to query pre-existing answers rather than performing iterative calculations during the actual simulation, thus resolving the contradiction between accuracy and computational speed.
Solution Approach 2:
The patent creates simplified surrogate models that replicate the complex thermodynamic calculation results. Instead of running full iterative thermodynamic calculations during reservoir simulation, the system uses these copied surrogate models (lookup tables containing pre-computed phase behavior data) to provide accurate phase stability and flash calculation results much faster, resolving the speed-accuracy tradeoff.
2Measurement precision
If comprehensive compositional data is processed through iterative thermodynamic calculations, then prediction accuracy is improved, but memory usage and computational resources worsen
Solution Approach 1:
The patent extracts only the essential thermodynamic properties and phase behavior data from comprehensive compositional analysis and stores them in compact lookup tables. By taking out only the critical information needed for reservoir simulation (phase stability indicators, flash calculation results) rather than processing all detailed compositional data iteratively, the system reduces memory usage while preserving prediction accuracy.
Solution Approach 2:
The patent performs comprehensive compositional analysis and thermodynamic property calculations in advance, storing the results in optimized lookup tables. This preliminary processing allows the actual reservoir simulation to use compact pre-computed data structures rather than maintaining and processing large volumes of raw compositional data in memory during simulation operations.
3Reliability
If traditional iterative methods are used for phase-split calculations, then thermodynamic accuracy is improved, but computational complexity and processing time worsen
Solution Approach 1:
The patent creates simplified surrogate models that copy the essential thermodynamic behavior from detailed equations of state. These surrogate models use pre-computed lookup tables containing phase-split calculation results, replacing complex iterative thermodynamic routines with direct table queries that maintain thermodynamic accuracy while dramatically reducing computational complexity.
Solution Approach 2:
The patent replaces the mechanical iterative calculation process with a data-driven lookup approach. Instead of repeatedly solving thermodynamic equations through iterative numerical methods, the system substitutes this mechanical computation with efficient data retrieval from pre-computed lookup tables, reducing computational complexity while preserving thermodynamic reliability.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
Technologies related to training machine-learning-based surrogate models for phase equilibria calculations are disclosed. In one implementation, an equation of state (EOS) for each of one or more regions of a reservoir is determined based on results of one or more pressure, volume, or temperature (PVT) experiments conducted on samples of downhole fluids obtained from one or more regions of the reservoir. Compositions of the samples of the downhole fluids are determined and spatially mapped based on interpolations between the one or more regions of the reservoir. One or more PVT experiments are simulated for the spatially mapped compositions of the downhole fluids using the determined EOS to create a compositional database of the reservoir. One or more machine-learning algorithms are trained using the compositional database, and the trained one or more machine-learning algorithms are used to predict phase stability and perform flash calculations for compositional reservoir simulation.