Using AI in Hydrodynamic Modeling of Oil and Gas Fields

Viktor Kotezhekov27 May 2023

Problem

An oil company has to forecast the production rate of every well, including a well planned at a new location. The inputs are scarce: wells are hundreds of meters to kilometers apart, and only a small share of them (possibly 5%) have studies. Geologists interpolate between wells by hand, then reservoir engineers run a grid simulator with millions of active cells and adjust the properties until the forecast matches history. A computation can take a day or a month, and a badly forecast reservoir costs more than well-level optimization (the speaker's example: about 3% effect).

Example from the article: given one field, 10 geologists produce 20 different models.

Solution in brief

Use ML where it is safe and keep physics where ML is weak. Build permeability maps from data with kernel regression plus a seismic CNN, tune the simulator with Bayesian optimization, and speed up forecasting with a Bayesian-regression ensemble of physics and ML models or with physics-informed networks (PINN, PINO).

Questions this case answers
  • How to build a fast ML surrogate for a slow physics simulator without producing non-physical forecasts
  • How to combine a physics-based forecast with ML forecasts so that one bad model does not spoil the result
  • How to combine physics-based and ML forecasts so that a single bad model does not ruin a well forecast
  • How to generate a property map from sparse measurements of different scale and reliability

Tools and methods

  • Nadaraya-Watson kernel regression — merges measurements of different scale and quality into a property map
  • 3D CNN on seismic attribute cubes — predicts permeability around wells
  • Bayesian optimization (Gaussian-process surrogate) — adapts the simulator, a black box
  • Bayesian linear regression ensemble, MCMC — weights physics-based and ML forecasts
  • PINN and Fourier neural operator (PINO) — fast approximation of the flow equation
Continue reading — it's free

Google or a one-time email code. No password.