Since the inception of multi-well pad development, operators have been searching for the optimal spacing and completion design. Historically, many operators have followed a pilot-and-develop method, investing a significant amount of up-front capital to test various spacing and completion designs, followed by a rigid implementation phase, using the plan from whichever pilot performed best. While this approach may be optimal for the pilot area, it may be sub-optimal elsewhere due to the changing geology, inter-well communication, and reservoir performance across a basin. Machine learning is a very robust solution to further optimize pilot studies across basins with thousands of existing wells where multiple variables are changing. For this paper, a few example machine learning models were trained using geologic, completion, and spacing parameters to predict production across the primary developed formations within the Midland Basin. These models were used to gain insight for three different studies: basin wide trends when inter-well spacing is widened, strategic infill well positioning to offset potential negative impacts from existing well depletion, and the optimal configuration of a new multi-zone development for an undrilled acreage.
For one study, the base case was eight wells each in the Lower Spraberry, Wolfcamp A, and Wolfcamp B stacked directly above each other. The model had a median absolute percent error (MAPE) of ~15% for a random held-out 20% of pads, a score that compares favorably to traditional manual reservoir engineering forecasting methods. Wells were then incrementally removed, and positioning was changed from stacking to staggering, production forecasted for each scenario to quantify inter-well communication, and results evaluated for the maximum value. The optimal case has two total fewer wells than the base case, with a stagger recommended across all zones. Although this case has a lower total recovery, the relatively low loss in oil justifies the capital savings of removing two wells.
The tremendous amount of data generated by hundreds of operators drilling thousands of pads across multi-zone plays has opened the door to the use of data-driven methods for unit-specific spacing optimization. The results from the study noted indicate that drilling fewer wells than the base case, and shifting those wells to a staggered development, results in a better return on investment. However, this workflow will yield different recommendations across the basin, a consequence of many changing factors including geology, inter-well communication, reservoir performance, spacing, and completion size. It is important to note that a data driven model must have a robust data set for the problem being evaluated. As an example, if the problem involves comparing stacking versus staggering, the data set must have a set of existing stacking and staggering examples to learn from. Thus, the approach of using machine learning to test several different combinations of spacing and completion designs can be repeated across a basin to find the most-economic, custom solution for each development unit.