O&G Well Stock Modeling Platform

News3 mins read

The problem

Oil and gas reservoirs are depleting at a frightening rate, projected to reach -10% by 2030. Maturing oilfields require better revitalization methods to alleviate or even negate this trend. How and to what extent can computer modeling complement well treatments for revitalization?

Forecasts to 2100 of production of four aggregations of classes of oil. Laherrère, Hall, Bentley 2022 (Curr Res in Env Sust), Total et al. https://doi.org/10.1016/j.crsust.2022.100174

Giant oil field decline rates and their influence on world oil production. Höök, Hirsch, Aleklett 2009 (Energy Policy), Uppsala University et al. https://doi.org/10.1016/j.enpol.2009.02.020

The solution

Our 3-level platform is a holistic approach of modeling different well treatments in a single AI-powered environment for the entire well stock — allows to achieve additional 10-20% increase in oil production, potentially bringing $35-70 billion of additional annual revenue to the whole O&G industry.

The product

Modules

  • Well treatments history analyser
  • Petrophysical & geomechanical constructor
  • Test injection analyzer
  • Formation damage
  • Radial drilling
  • Hydraulic fracturing (in dev)
  • Acid frac
  • Matrix acidizing
  • Water shutoff
  • Profile control

Functions

  • Well stock parallel modeling
  • Automated well ranking
  • Well analytics & planning
  • Well treatment optimization
  • Library of reagents & treatment technologies
  • Library of best practices (in dev)
  • Corporate database integration
  • Regular updates Technical support

Benefits

  • Production increase
  • Incidents risk reduction
  • Treatments cost saving
  • Technological superiority
  • Precise tuning of injection parameters
  • Optimal reagents selection Preliminary performance evaluation
  • Project portfolio management
  • Detailed reporting

Customer cases

  • Created detailed acid treatment designs: >1000
  • Created template acid treatment designs: >500,000 (the whole well stock), 2500 best wells identified
  • Average production rate increased by 12%
  • The time spent on modeling decreased by 50%. Field and hydrodynamic research costs reduced by 30%
  • High convergence of simulation results with actual data: >90% for oil production rate increase and >87% for skin
  • Validated new formation damage module in the acid treatment modeling of the injection well stock
  • Created acid treatment designs for 15 wells
  • High convergence of stimulation results with actual data: >90% for oil production rate increase
  • Validated simulator performance in low-temperature sites of complex wells (3 wells with dolomites)
  • High convergence with actual data: >98% for oil production rate increase and >94% for skin

AI Component

Limitations of Deterministic Numerical Models at High Workload

In our Well Stock Manager, the computational costs of deterministic methods may become prohibitively high, especially in cases when multiple well treatment types are being modeled for large well stocks with a considerable number of design iterations for each well+treatment.

Traditional simulations using finite difference method:

  • Require a lot of computational time
  • Scale poorly with problem size and dimensionality

Incorporating graph neural networks (GNN):

  • Allows for fast inference once trained
  • Significantly reduces computational time while preserving the accuracy

Learning Mesh-Based Simulation with Graph Networks Pfaff, Fortunato, Sanchez-Gonzalez, Battaglia 2021 (ICLR), DeepMind

Mesh-Based Flow Simulation using Graph Neural Networks

Irregular meshes (right) are far better suited to model general, non-rectangular geometries accurately compared to regular grids (left)

A dynamic mesh-based simulation, mapping fluid velocity over time in the presence of an obstacle

Overview of the GNN model training process

For the high-volume tasks we are to replace our deterministic numerical models with GNN (where graphs represent formation geometry and structure), trained on the results of our regular models involving systems of linear equations. While the accuracy (convergence) of the GNN remains high, the computational speed is 1-2 orders of magnitude faster than the simulation on which it is trained.

Learning Mesh-Based Flow Simulations on Graph Networks Rayan Kanfar 2022 (Medium), Stanford University, Saudi Aramco

For more details, contact us.

SHARE: