Use of Numerical Simulation Enhanced by Machine Learning Techniques to Optimize Chemical EOR Application

Abstract

Leveraging the recent developments in the Machine Learning (ML) technology, the objective of this work was to use Artificial Neural Networks to build proxy models to classical reservoir simulation tools for two distinct chemical EOR applications. Once built and calibrated (trained), these ML-based proxy models were used to efficiently identify optimal scenarios to be further considered in the corresponding EOR developments, therefore demonstrating how these techniques can complement classical tools to enhance the decision-making process. Two numerical simulation models were built and calibrated to reproduce lab-measured data from a real surfactant-polymer coreflood experiment (Application #1) or historical data from a real oilfield (Application #2). Different scenarios were then simulated: Application #1: various sequences of injection were explored (chemical concentrations and slug sizes)Application #2: different surfactant-polymer injection configurations were investigated on a large-scale multi-pattern configuration Simulated outputs were used to train Artificial Neural Network models, which were checked for their predictivity on unseen data. These ML-based proxy models were finally used to rapidly identify other optimal scenarios for each application based on several economic indicators. For the first application, numerical model calibration was obtained using one real coreflood experiment: measured pressure signal was well reproduced by the simulator, as well as oil and surfactant production. Several numerical simulations were then performed to evaluate the oil recovery from different injection sequences. Both surfactant and polymer concentrations were varied as well as the slug's durations. For the second application, a history-matched sector model representing the current status of a real oilfield after several years of waterflooding was used. Several scenarios were simulated to evaluate oil recovery associated with distinct sequences of EOR injection consisting of surfactant and polymer agents of various slug volumes and concentrations. Using a train/test split approach, 80% of the simulations were used to train one Artificial Neural Network for each application. The remaining simulations (20%), used as blind tests, confirmed the predictivity of the trained models on unseen data. The ANN models were finally used to predict outcomes from new scenarios not investigated by numerical simulation. This enabled us to identify optimal scenarios with regards to classical economic indicators. These scenarios were then numerically simulated to confirm the predictions from the ANN models, therefore validating the whole approach. This work illustrates how modern machine learning techniques such as Artificial Neural Networks can be used to enhance the numerical simulation tool while solving specific optimization problems (here related to chemical EOR application). Such techniques, becoming increasingly accessible thanks to open-source programming languages, provide a powerful lever to become more efficient when using reservoir simulators as a tool to guide decision-making processes in the oil industry.