Novel Transfer Learning Workflow Allowing Flexible Well Configurations for Physics-aware Deep-learning based Proxy Reservoir Simulation Models

Transfer Learning techniques have been employed in a variety of simulation and optimization problems across multiple industries to endow deep machine learning models with predictive capabilities for changing boundary conditions (or targets in labeled data) or in cases with limited data. Application of Transfer Learning to Hybrid (Physics/Data based) Deep learning-based proxy reservoir simulation can lead to an efficient and reliable tool for reservoir engineers. In this paper, we propose the application of transfer learning to a physics-aware deep-learning based proxy reservoir simulator to enable prediction of well outputs and state variables in a dynamic reservoir simulation workflow. We apply transfer learning techniques to extend the capabilities of our recently developed deep-learning based proxy reservoir simulator titled Embedded to Control and Observe (E2CO) to allow for fast and robust prediction of well production from a pre-trained model with flexible well configuration. In this research, the original, physics-informed based deep neural network proxy model of an existing reservoir is retrained at key layers within the structure of the network. Next, various adjustments are made in the properties of the original model such as well placement, scheduling and controls. Utilizing only a fraction of training material required for the original model, we then quickly re-trained existing models with altered properties to reliably predict the output from different well configurations. Our testing displayed good matches on both oil and water production after retraining an existing 4 Injector and 5 producer model with adjusted producer well locations and numbers. Re-training said model with adjusted well positions resulted in a 2% error on cumulative oil and water production matches as compared with numerical solutions, and a 6% error with the addition of new wells. The methodology proposed in this research is a step towards the development of a fast and robust replica of a high-fidelity model which can be employed throughout the complete reservoir management lifecycle for the prediction of various reservoir layouts. The system itself is agnostic to the type of reservoir problem at hand and is applicable to oil production and carbon sequestration problems alike. Such a method provides a fast alternative to numerical simulation output in the development of a reservoir. Significance of Paper Shows applicability of transfer learning techniques in creation of proxy models for reservoirs as fast alternatives to numerical simulations. Moreover, it leads to computational power saving by allowing us to reuse completed computations in the development of reservoirs.