Analyzing Midland Basin Well Performance and Future Outlook with Machine Learning

The Midland Basin, now producing over 2 million barrels a day from horizontal wells, has become a cornerstone of Lower 48 oil production. However, per-well productivity declines over the last two years have clouded the basin’s outlook, with potential causes including rock quality degradation, inefficiencies from long laterals, and parent-child interactions. This study utilizes advanced machine learning to analyze these factors and forecast remaining inventories. Our findings challenge the narrative of production decline, attributing it to temporarily elevated outputs in 2020-2021 due to operator high grading. We compare remaining inventory to already-developed and find at least a decade of inventory life remaining.

First, we built a dataset from public and proprietary sources, encompassing completions designs, subsurface properties, directional surveys, and production data. We calculated parent-child relationships and interwell spacing alongside landing zones for each well. Subsequently, we trained a machine learning model to forecast a three-year time series of oil, gas, and water production at monthly increments. This model included a causal inference component for accurately forecasting tight spacing and parent-child scenarios. We utilized SHAP Values (SHapley Additive exPlanations) to analyze rock quality and performance trends over time in the historical dataset. Finally, we forecasted remaining inventory, employing a data-driven approach for spacing and development assumptions.

We find the Midland Basin’s production decline mainly resulted from high productivity in 2020-2021 due to operator high grading, with 2023’s average rock quality drilled similar to that of 2019. In aggregate, about 50% of well locations in the basin have been drilled, though the drilled fraction is higher in the best areas of the basin. Exploring above and below previously developed zones offers potential to extend the basin’s inventory, although these infills typically yield lower performance due to parent-child communication. Using current activity levels to extrapolate a future development schedule, oil production may dip slightly in 2024 but is projected to stabilize around 2 million barrels a day through the mid-2030s.

This study highlights the substantial remaining potential in the Midland Basin, putting the recent productivity rise and fall in the context of commodity prices and operator decisions. With around 10 years of remaining inventory in primary zones, the basin is well positioned to continue production levels into the 2030s. A significant novelty of our approach is the causal inference machine learning models’ ability to forecast parent-child and infill developments more accurately, which every year becomes a larger share of the remaining inventory.