In previous posts, we have examined underlying trends in horizontal well datasets, the implications of these datasets on data-driven modeling efforts, and Novi’s approach to dealing with biased data. In part 2, we created a data-driven model using 11,176 wells from the Delaware Basin and features representing geology, completion intensity, spacing, and timing. In part 3, we discussed how to make a de-biased model that identifies sub-trends which make physical sense in response to spacing and timing changes.
Now, we will create a causal model and compare its results to the standard associative model of part 2. This will involve investigating both the overall variable response to spacing and timing, as well as creation of planned well scenarios with different spacing packages. In part 2, we discussed the model variables and definition of spacing neighbors. We used the farther distances shown here to provide interwell spacing information to the model. These are the only 2 spacing features that this particular model can see.
Comparing SHAP Responses: Standard vs Causal Models
In part 2, we also looked at the dataset trends comparing overall well performance to these distance variables, as well as the SHAP trends generated by the model. We observed that the model does identify an increased production trend with wider spacing, but by a magnitude of less than 10% when taking a well from a 440ft spacing to unbounded. Now, we can compare how a standard pure data-drive model compares to a causal model which gives special treatment to spacing and timing variables, as described in part 3.
Figure 2: A) SHAP response to stagger farther neighbor in standard model. B) SHAP response to lateral neighbor in standard model. Directionally, the model learns the correct relationship between spacing and production, but the magnitude of penalty for tight spacing is insufficient. C) SHAP response to stagger farther neighbor with causal modeling approach. D) SHAP response to lateral farther neighbor with causal modeling approach. The causal model successfully buckets wells into similar geological and completions characteristics, and finds a much steeper, nonlinear relationship between well performance and spacing.Testing Spacing Scenarios in the Delaware Basin
Now that we have observed the overall difference in model response to interwell spacing, we can test a sample pad in the Delaware Basin with different spacing packages. Figure 3 shows the sample pad location in Lea County. Figure 4 shows the different spacing scenarios we ran through both the standard and causal models. The scenarios included were (4, 6, 8, 10) wells per section per zone in the First and Second Bone Spring Sand, and the Wolfcamp A and XY. For interior wells, this corresponds to lateral spacing varying from 528 feet to 1320 feet. Stagger well distances are much tighter in the Wolfcamp targets than the Bone Spring targets due to the zone thicknesses in this area.
Figure 3: Location of test pad in the Delaware Basin, Lea County. There is currently no existing development at this location and the nearest neighbors are more than a mile away. This gives a controlled experiment to explore model responses to interwell spacing.
Figure 4: Winerack scenarios for Delaware Basin test pad. Increasing the well density from 4 to 10 wells per section per zone should result in significant production forecast declines if the model has been properly conditioned and debiased.In Figure 5, we explore the results of these pad spacing scenarios at 2500 lbs/foot of proppant and 2500 gals/ft of fluid injected. In the standard model (Figure 5D), the 10 wells per section scenario is forecasted to produce within 97% of the 4 wells per section scenario.This result is caused by the sampling bias in the dataset. As the wells move closer together, the model is forced to seek analogs in high rock quality areas. Figure 5C shows the same result on a causal model. The Wolcamp zones show degradation on this pad between 15% and 20%, and Bone Spring zones degrading by10%. The magnitudes and relative declines will change based on completion size and geological area, but the difference between standard and causal model results is directionally consistent.
Figure 5: Forecasted 2-Year cumulative oil production for different spacing scenarios on the Delaware Test Pad location. In the standard associative model, the production decline is essentially flat as spacing is tightened. The standard model is not responsive to spacing. In the causal model result, there are monotonic declines in production as the spacing patterns are tightened. Panels A) and B) show the per well average cumulative production by zone and spacing scenario. A) shows the causal model result and B) shows the standard model result. Panels C) and D) show the spacing degradation relative to the 4 WPS scenario. The Wolfcamp zones have a steeper degradation due the proximity between the Wolfcamp a and Wolfcamp XY. As the spacing packages are tightened, Wolfcamp wells experience degradation from both lateral in zone neighbors and the stagger neighbor in the other Wolfcamp zone.Conclusion
The causal model provides a more accurate understanding of well spacing, revealing production declines that standard models overlook due to bias. By accounting for geology and completion strategies, the causal model helps operators make better-informed decisions, optimizing well performance and reducing risks in field development. This approach is a key step toward improving well forecasting and planning in complex scenarios
Next Steps: Different Methods of Calculating Spacing Information
In the next post, we will explore the different methods of calculating spacing information.
Key Takeaways
- A Delaware Basin model built on 11,176 wells used only two spacing inputs, the stagger farther neighbor distance and lateral farther neighbor distance.
- SHAP in the standard associative model showed less than a 10% uplift when widening spacing from 440 ft to unbounded.
- SHAP responses in the causal model showed a steeper, nonlinear relationship between well performance and interwell spacing than the standard model.
- A Lea County Delaware Basin test pad with nearest neighbors more than a mile away enabled controlled spacing scenario comparisons.
- Winerack scenarios tested 4, 6, 8, and 10 wells per section per zone across First/Second Bone Spring Sand and Wolfcamp A/XY, corresponding to 528–1320 ft lateral spacing.
- At 2,500 lbs/ft proppant and 2,500 gal/ft fluid, the standard model forecast the 10 WPS case within 97% of the 4 WPS case.
- At the same completion intensity, the causal model forecast 15%–20% degradation in Wolfcamp zones and ~10% degradation in Bone Spring zones from 4 to 10 WPS.
