Machine learning (ML) techniques are emerging as transformative tools for field development planning by offering rapid and accurate predictions of the subsurface. Hybrid conventional-ML workflows enable engineers to swiftly and systematically screen for optimum solutions in the nearly infinite design space.

To evaluate their effectiveness, this paper presents an end-to-end case study where Fourier neural operators (FNOs) are employed within a CO2 well placement workflow. The workflow combines data generation with a commercial simulator, network training, and optimization. We assess the performance of FNOs by quantifying their accuracy versus training data quantity. The simulator and ML models are then coupled with a genetic optimizer for combinatorial well placement. We analyze how FNO accuracy impacts the optimization quality and the effectiveness of the FNO-surrogate and simulator-only workflows.

The FNO-enabled workflow was found to provide superior well placement results, faster total workflow time and lower compute costs under all conditions tested.

Keywords:
geologist, fluid dynamics, deep learning, subsurface storage, neural network, artificial intelligence, climate change, simulator-only workflow, partial differential equation, workflow
Subjects:
Reservoir Fluid Dynamics, Improved and Enhanced Recovery, Reservoir Simulation, Storage Reservoir Engineering, Environment, Asset and Portfolio Management, Information Management and Systems, Flow in porous media, CO2 capture and sequestration, Climate change
Copyright 2025, Society of Petroleum Engineers DOI 10.2118/224159-MS
You can access this article if you purchase or spend a download.