Forecasting oil recovery under polymer enhanced oil recovery technique using machine learning

• Among the first applications of time series ML to forecast cumulative oil recovery • Highly accurate forecasts without requiring detailed reservoir characterization • Polymer flooding increased cumulative oil recovery by 23.8% over waterflooding • 60/40 train–validation split gave the least errors (MAPE∼0.784%, RMSE∼0.236) • Potent tool to assess the feasibility and success of Enhanced Oil Recovery projects Enhanced Oil Recovery (EOR) remains a crucial technique for sustaining hydrocarbon production amid rising global energy demand. However, the complex fluid-rock interactions within a reservoir continue to pose significant challenges in accurately forecasting oil recovery under various EOR techniques. This study presents a novel application of Time Series Machine Learning (TSML) in the development of a data-driven framework to forecast cumulative oil recovery under laboratory-scale EOR operations. Recovery Factor (RF) data was sourced from a single polymer-based coreflooding experiment, where the cumulative recovery reached 49.05% of the Original Oil In Place (OOIP). This total comprised an initial recovery of 25% from the waterflood and an additional recovery of 24.05% from the subsequent polymer EOR stage. An Auto Regressive Integrated Moving Average (ARIMA) model was developed using the oil RF data to forecast the future recovery trajectory, eliminating the need for detailed reservoir characterization. Under a 60/40 train-test split, the optimized ARIMA (1,0,2) model projected a terminal RF value of 23.62%, demonstrating a marginal deviation of only 0.43% RF relative to the experimental baseline. This predictive accuracy was quantified by an RMSE of 0.236 and MAPE of 0.784%. The present work facilitates rapid EOR feasibility screening by reducing the injection duration to 60%, an efficiency gain that translates from hours in laboratory to months in the field. While not a recovery-enhancing mechanism itself, the selected forecasting method provides a data-efficient and lightweight supplement to conventional simulators for objective, scalable pilot and field-scale evaluation.