Old Wells, New Math: Predicting Refrac Success

Machine learning tools help US shale operators predict which legacy wells still justify another frac

1 Aug 2025

Refracturing is back in the conversation across mature U.S. shale plays. This time, the story is less about a wave of new jobs and more about sharper decision making. Operators are turning to machine learning to help answer a stubborn question: which old horizontals still deserve another frac, and which ones do not.

The appeal is straightforward. Refracs carry risk. Years of production change stress, drain pressure, and increase interference, all of which can derail results. Traditional screening methods often struggle to capture those effects. Data-driven workflows are now stepping in to narrow the unknowns.

Recent technical work points to a shift from hindsight to prediction. At URTeC 2025 in Houston, one study outlined a workflow that combined diagnostics, production history, and integrated simulation to blind-predict refrac outcomes across datasets in the Bakken and Midland Basin. The goal was not to explain past wins, but to test whether performance could be forecast before pumping another stage.

Industry coverage reflects the same direction. A Journal of Petroleum Technology feature on refracturing technologies described machine learning tools used to explore completion optimization, based on discussions from URTeC sessions focused on AI and emerging methods.

Beyond single papers, the trend is becoming more formal. Conference listings from ADIPEC describe physics-inspired, data-driven frameworks designed to flag underperforming wells that may still hold value. These systems aim to bring consistency to candidate selection, rather than relying on ad hoc judgment.

Service companies are supplying much of the plumbing. Halliburton, SLB, and Baker Hughes all promote AI-enabled subsurface and production analytics platforms. While not branded as refrac tools, they provide the building blocks researchers and operators adapt when data quality allows.

Adoption remains uneven. Machine learning models are sensitive to missing records and local geology, a common issue for older wells. Still, progress is steady. As datasets improve and models are paired with diagnostics and engineering judgment, refracturing decisions are becoming more structured, more comparable, and less reliant on trial and error.