A Novel Data-Driven Approach for Capacitance Resistance Models in Multi-Phase Flow Systems

Abstract Capacitance Resistance Models (CRM) are increasingly recognized for their role in managing waterflooding and offering a streamlined subsurface representation for operational optimization. Since their introduction two decades ago, CRMs have evolved to address multi-phase flow and, with the integration of machine learning algorithms, have been extended to applications in gas, gas-alternating-water injection, and other enhanced oil recovery mechanisms. In this paper, we introduce a novel data-driven formulation that reduces the number of optimization parameters by 50% compared to traditional CRM approaches. This advancement not only enhances deployment efficiency in fields with numerous wells but also significantly accelerates CRM runtime. We validated our new formulation using several reproducible subsurface numerical simulation models that have been previously employed in CRM studies. Our validation also included tests on more compressible fluid systems (such as gas injection) and scenarios with significant changes in oil-water mobility. CRMs provide critical insights into well-to-well correlations in flooded reservoirs and are typically used for production optimization. However, this often requires the separation of production phases into water, gas, and hydrocarbon liquids, which is typically achieved using additional analytical or data-driven models in conjunction with CRM. In response, we present a novel method that integrates CRM with a pattern balancing algorithm, enabling optimization based on multiple objectives such as reservoir fill-up, sweep efficiency, and water injection effectiveness. Our extended formulation not only improves the accuracy of CRM models—evidenced by increases in R2 values from 0.6 to 0.95 in some cases—but also provides more precise subsurface insights, reducing the risk of misinterpreting reservoir trends as false well correlations. Finally, we explore the complementary role of production data-driven models alongside more detailed numerical reservoir simulation models.