Improving deep learning performance for predicting large-scale geological [Formula: see text] sequestration modeling...

Area of Science:

• Computational geoscience
• Artificial intelligence in subsurface modeling
• Porous media physics

Background:

• Physics-based reservoir simulation is computationally expensive for large-scale 3D heterogeneous porous media.
• Predicting temporal-spatial patterns of state variables like pressure requires significant computational resources.

Purpose of the Study:

• To develop an efficient deep learning (DL) workflow for predicting pressure evolution in large-scale 3D heterogeneous porous media.
• To accelerate reservoir simulation using DL by reducing computational expense and improving prediction efficiency.

Main Methods:

• Developed an efficient feature coarsening technique for DL training and prediction at a coarse scale.
• Utilized spatial interpolation to recover fine-scale resolution from coarse-scale DL predictions.
• Validated the DL approach against physics-based simulation data for a 3D geologic sequestration reservoir model.

Main Results:

• Feature coarsening decreased training time by [Formula: see text] and memory consumption by [Formula: see text], while maintaining an average temporal error of [Formula: see text].
• The DL workflow achieved a 1406x speedup compared to traditional physics-based simulations.
• Demonstrated significant improvements in training and prediction efficiency for large-scale heterogeneous reservoir models.

Conclusions:

• The proposed DL workflow offers a highly efficient alternative to conventional physics-based simulations for reservoir pressure prediction.
• This method significantly enhances computational efficiency, enabling faster history matching and reservoir optimization for closed-loop reservoir management.