Interpretable Machine Learning for Estimated Ultimate Recovery (EUR) Prediction for the Bakken Tight Oil Play, Williston Basin, USA

ABSTRACT Estimated Ultimate Recovery (EUR), often calculated from production decline curve analysis (DCA), is a key indicator of long‐term well performance in petroleum reservoirs, including unconventional plays such as the Bakken Formation in the Williston Basin in the northern United States, which is the objective of this study. Long‐term production data are necessary for estimating EUR. However, for fracture‐stimulated wells, it is necessary to first isolate restimulated wells from the large set of well data used for EUR estimation. This study presents an application of interpretable machine learning (ML) workflow to the Bakken Formation, aiming for (1) identification of restimulated wells from a set of 7830 horizontal wells, and (2) understanding the contribution of various geospatial, drilling, and completion factors to EUR. Shut‐in months are first removed with a one‐sided Hampel filter. Restimulated wells are then identified by a random forest (RF) classifier built on 10 features derived from a nonparametric isotonic‐regression baseline and trained on reported production and completion data. The results revealed that the classifier achieves a cross‐validated area under the receiver operating characteristic curve (ROC‐AUC) of 0.877 and flags 1140 wells, including 672 candidate restimulated wells that have no restimulation record in FracFocus. For the 4510 filtered wells, EUR is obtained from a 3‐year modified hyperbolic DCA, which reproduces observed cumulative production with an R 2 of 0.9 when its 3‐year forecast is compared against the actual post‐3‐year production of the wells. EUR per hydraulic fracturing spacing (EUR s ) is then predicted from nine parameters derived from including geospatial (surface longitude and latitude), drilling (true vertical depth and lateral length of each well), completion (number of hydraulic fracturing stages, water, and proppant mass), and production (first‐month production and first‐month gas‐to‐oil ratio) data. The ML prediction was achieved using the Super Learner algorithm, yielding R 2 = 0.758 on the test dataset. SHAP analysis was employed to interpret model behavior and identified that the number of hydraulic fracturing stages, surface longitude, and water index (gallons per hydraulic fracturing spacing) were the three features most strongly associated with EUR s . This work presents an integrated and reproducible workflow that removes restimulated wells from production history and leverages the impact of drilling and completion design, early production data, and geospatial relationships on EUR estimation using interpretable ML.