Metamodelling colossal plans

Generative Models:

• Gas injection
• Aquifer
• Tables (PVT \ RPP)

• Better porosity, permeability, initial distributions

• Geometry ?? (Ivan Makhotin)
• Trajectories of wells ??
• Hydrodynamical fracturing

Unconventional History-Matching

Production optimization
(Optima will not work properly):

• Ensembling of neural networks
• Risk-aware optimization
• Entropy-seeking exploration

Aquifer options:

• Check if substitutable with injection wells
• Check if NN can work properly with it in a given architecture

Super-resolution:

• Simple upscaling (several types)
• Downscaling with NN (probably GAN or other generative network)
• Exact shape match does not matter!
• Combining forecasts for different upscalings

Metamodelling base in the framework:

• Grid search for parameters and attributes
• Simple block-based architectures (combine layers in nn.Sequential and go)
• Platforms for non-linear geometry and graphs in a single solution
• Micro-batching
• Use Metaflow (Vadim Kuzmin)

tNavigator-like API

• Run tNav in cycle
• Make predictions for whole Fields, not samples

Control:

• Control component in Field
• Group control in the metamodel
• Simple control with inequalities in the metamodel
• Iterative check for non-negative rates in the metamodel

Rates:

• Is it working on corner point?
• Differentiable loss with rates?

Tables:

• Table representations for metamodel:
  • PCA
  • NN-based
  • Physics-aware??
• Regions?

Hydrodynamical fracturing:

• Representation in the metamodel
• Problems with symmetric influence
• May be, it is enough to consider improvement of production, not exact representation of states.