Graph Partitioning for Parallel Simulation Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current graph partitioning algorithms for parallel simulation models in hydrocarbon reservoirs are inefficient, particularly when dealing with large ranges of real weights, as they fail to preserve critical grid blocks or physical properties within subdomains, leading to slow convergence or failure in simulations.

Innovation Solution

A method that maps real-valued connection weights to a narrow range of integer weights using a cumulative distribution function and scaling techniques, ensuring that wells and other reservoir features remain within a single subdomain during partitioning, thereby preventing partitions from crossing near well regions or geologic features.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If graph partitioning algorithms are used to divide simulation models into subdomains for parallel computing, then computational efficiency and simulation turnaround time are improved, but partitioning quality deteriorates when dealing with large ranges of real weights, causing wells and critical grid blocks to be separated across subdomains

Engineering Contradiction:
Improvesimulation turnaround timeVSAvoidpartitioning quality
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies parameter changes by transforming the weight values from a large range of real numbers to a limited range of integer values. This transformation is achieved through a mapping function that preserves the relative relationships between weights while fitting them within the constraints required by partitioning algorithms, thereby improving partitioning quality without sacrificing computational efficiency

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces an intermediary transformation layer between the original real weights and the partitioning algorithm. This intermediary mapping function serves as a mediator that converts the continuous weight values into discrete integer values, enabling the partitioning algorithm to effectively handle the weight information while maintaining the critical structure needed to preserve wells and grid blocks

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If partitioning algorithms assign weights to graph connections to preserve critical features, then partitioning quality improves, but computational complexity increases due to the mathematical analysis required

Engineering Contradiction:
Improvepartitioning qualityVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by performing the weight transformation and mapping before the actual partitioning process. By pre-processing the weight values to convert them into suitable integer formats, the system simplifies the subsequent partitioning computation, reducing the overall computational complexity while maintaining high partitioning quality

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2534605B1Method and system for partitioning parallel simulation models
Publication Date: 2020.06.17 EXXONMOBIL UPSTREAM RESEARCH COMPANY(US)
  • EP2534605B1 patent drawingFigure 1
  • EP2534605B1 patent drawingFigure 2
  • EP2534605B1 patent drawingFigure 3

AI summary

A method is presented for partitioning a simulation model into a plurality of subdomains that may each be assigned to one of a plurality of processors. The method includes creating a representation of a topology graph of a simulation model in a tangible, computer readable medium. The topology graph includes a plurality of computational elements and a plurality of connections between those elements. Each of the plurality of connections is weighted to create a plurality of weights, and each of the plurality of weights is scaled. Optionally, the weights can be mapped to different interval of values. Based on the weights information the topology graph is partitioned into two or more subdomains, wherein a partition boundary follows a local topographical minimum in the topology graph. A subdomain is assigned to each of the plurality of processors.