Transfer Learning for Variable Well Locations and Permeability Distributions in Physics-Aware Deep Learning Reservoir Simulation Proxy Models

Abstract Transfer Learning techniques have been utilized in a variety of simulation and optimization problems, enhancing deep machine learning models with the ability converge under changing boundary conditions, varying targets in labeled data, and scenarios with limited data availability. This research proposes applying transfer learning to a physics-aware deep-learning-based proxy reservoir simulator titled Embed to Control and Observe (E2CO). In this study, the original physics-informed deep neural network proxy model of an existing reservoir is allowed to retrain key layers within the network's structure. Various properties of the data are adjusted, such as well placement, permeability fields, scheduling, and controls. By utilizing a fraction of the training material required for the original model, we quickly re-trained existing models with altered properties. Our tests showed excellent matches for cases of reservoirs with varying well locations for both oil and water production after retraining a model with 4 injectors and 5 producers, and with the addition of new wells. Retraining the model with modified well positions resulted in an average of 3% error on cumulative oil production, and 5% on cumulative water production matches compared to numerical solutions. Modifying permeability fields yielded mixed results. This methodology is an advancement towards creating a fast and robust replica of high-fidelity models, enabling rapid, low-cost, computational predictions of various reservoir layouts. The system is versatile and non-intrusive, applicable to oil production and carbon sequestration problems alike. This method offers a fast proxy to complement traditional numerical simulation, providing a significant boost in efficiency and flexibility.