From Prediction to Decision: A Multihorizon Completion Optimization Framework for the Bakken Field

In many applications, machine learning (ML)-based production forecasting has been used primarily as a prediction tool rather than being embedded within an end-to-end optimization workflow that directly informs completion design or economic decision-making. To address this gap, we develop an integrated three-stage framework—comprising feature engineering, surrogate modeling, and optimization—and test it on historical data from the Bakken. Specifically, we introduce a spatial-temporal feature engineering method that captures interwell interference and parent/child relationships without relying on raw coordinates. We then use extreme gradient boosting (XGBoost) as a multihorizon surrogate to predict cumulative production across 3-month to 24-month intervals, coupled with a constraint-based genetic algorithm (GA) to optimize completion designs for either maximum production or net present value (NPV). The XGBoost model achieves a coefficient of determination (R²) above 0.9 on the training set and above 0.6 for the 12-month to 24-month horizons on the testing set. According to the explainability study, the two factors with the largest consistent impacts on production performance are the pressure gradient and cumulative oil production of the parent well. Applying this framework to testing wells demonstrates significant potential, achieving an average production increase of roughly 84,000 barrels and an NPV uplift of USD 2.47 million per test well. These results validate that integrating predictive modeling within a constrained optimization workflow yields superior, economically viable completion strategies, with the optimizer providing a 3.5:1 marginal return on additional completion costs.