Unconventional operators produced tremendous performance improvements over the last decade, with per-well oil production more than doubling in the Bakken and tripling in the Permian. This dramatic productivity improvement has been the result of longer laterals, more intense completions, and drilling of higher-quality acreage and landing zones. Understanding the relative contribution of each of these factors will help answer whether productivity will continue to improve, stagnate, or even regress. To investigate this topic, we built machine learning models in the Bakken, Eagle Ford, Midland Wolfberry, and Delaware Wolfbone plays. We then use these models to analyze well performance and the contributions from subsurface rock quality and well designs across each play. We aggregated production for horizontal wells with known completions designs, surface-hole locations, and bottom-hole locations across the basins listed above. We then trained a series of models to forecast a three-year stream of production using interpreted subsurface, publicly available completions parameters, and well spacing. After the initial model build, we trained a surrogate model to produce explanations for the model forecasts, specifically using SHAP values (SHapley Additive explanation), which isolate the contribution of each training variable on the model forecast. We examined average 1-year cumulative production values, analyzing both per-foot and absolute values, removing standalone parent wells. Longer laterals have had the largest impact on production, up to ~45% in the Delaware Basin, though the impact is muted in the Bakken where operators converged on 2 mile laterals much earlier due to state drilling unit regulations and favorable geology. Larger completions contributed on average a 20-30% increase, though their impact plateaued beginning in 2017-2018. Geological high-grading shows largest impact of ~50% in the Midland Basin, where activity shifted to the higher-quality northern portion of the basin from 2012-2018, but only ~10% in other plays. The results of this study are consistent with the hypothesis that future performance gains will be minor compared to those of the past because the key innovations of longer laterals and upsized completions have been exhausted. Though our models have not yet seen a significant dip in average rock quality drilled across a basin, that will eventually become a headwind due to inventory depletion. Machine learning models combined with explainability datasets provide a powerful tool to quantify the impact of operator design choices, analyze basin-wide trends, and inform expectations for future developments.

You can access this article if you purchase or spend a download.