The confidence game: optimizing completions for XEC DUC wells in the Delaware using model confidence as a proxy for risk

Introduction: Problem we are trying to solve; methodology we are applying to solve it

We have written a couple of posts about optimizing completions on DUC (Drilled but UnCompleted) well inventory given the extreme constraints on capital that Operators are facing. In our first post, we focused on QEP’s inventory in the Midland, and then in a second post, NBL’s inventory on the Delaware side. In both of those cases, the DUC inventory was relatively concentrated in a couple of development units.

We wanted to take a look at DUCs that were a bit more spread out, so we took at look at XEC, which has a vast acreage position in the Delaware spanning across a large swatch of the northern Delaware basin. The vastness of their position is due to some degree to their acquisition of Resolute Energy, which closed in March of 2019.

XEC has 50+ DUCs according to Enverus, spread across a very wide swath of the Delaware, some in the super deep column on the Texas/New Mexico border and running south from there, and some more out to the western fringe of the basin. Given this spread, the interaction between spacing, stimulation design, and subsurface makes the question of optimization a double edged sword – one must take into account how confident any model is in its predictions as you move from the thickest column in the core to the thinnest in the fringe, while at the same time understanding which completion designs applied to the DUCs are going to drive the best possible returns in the short term while minimally compromising EURs.

In this machine learning use case, we will demonstrate the utility of Novi’s implementation of Shapley values to understand interactions between well spacing, stimulation and subsurface data. Shapley values are based on the work of mathematician and economist Lloyd Shapley, which have their roots in game theory. Shapley’s work was later adapted for large scale adoption in the machine learning world by a couple of University of Washington students . Shapley data will be applied to help determine which completion design choices across Cimarex’s (XEC) DUC position in the Delaware basin would drive the most accretive value at a variety of oil strip price decks.

In addition, and perhaps more important when oil companies are operating on negative margins with a strip price in the $20 range, we will utilize Novi’s confidence intervals to guide recommendations on completion designs across XECs DUC well inventory, so the confidence the Novi model has in any given prediction is clear when evaluating the financial implications of each decision.

Where is XEC’s DUC inventory?

We loaded DUC locations for Cimarex in the Delaware basin into Forecast Engine and Spotfire. There are 50+ DUCs split between several units across their position. We are going to run these DUCs at 6 different completion designs to see if we can gain insights on which designs should be run at each location.

Historically, Cimarex has employed one completion design on their inventory:

• XEC 2,500 lbs/ft (0.8 Fluid): 2,500 lbs/ft of proppant and 2,000 gals/ft of fluid

We will take this completion design and flex it on the fluid ratio going up to what offset operators are doing in the area at a 1.0 & 1.2 fluid ratio. We will also flex the proppant by increasing and decreasing by 500 lbs/ft off of Cimarex’s design.

• XEC 1,500 lbs/ft (0.8 Fluid) : 1,500 lbs/ft of proppant and 1,200 gals/ft of fluid
• XEC 2,000 lbs/ft (0.8 Fluid) : 2,000 lbs/ft of proppant and 1,600 gals/ft of fluid
• XEC 2,500 lbs/ft (1.0 Fluid) : 2,500 lbs/ft of proppant and 2,500 gals/ft of fluid
• XEC 2,500 lbs/ft (1.2 Fluid) : 2,500 lbs/ft of proppant and 3,000 gals/ft of fluid
• XEC 3,000 lbs/ft (0.8 Fluid) : 3,000 lbs/ft of proppant and 2,400 gals/ft of fluid

Maximizing 2 year cums with completions optimization

After running XEC’s DUC inventory through Novi Forecast Engine at the 6 different completion designs from above, we can begin to gain insight into what is driving the production of these wells. Comparing the 2-Yr Oil cumulative volumes across the 6 designs, the Novi model predicted that the larger proppant loading increased the volumes; not too surprising. More interesting that that – increasing the fluid ratio provides minimal impact on the wells performance, as you can see by the flatness of the middle part of curve below. Doubling the job from a 1,500 lbs/ft to 3,000 lbs/ft analysis increases 2-Year oil cum by 6.4 bbls/ft or a ~21% increase. These are valuable insights but only enough to gain a general view of the inventory.

Once we dive into the Novi Shapley data generated for each well at 30 day prediction intervals we can begin to see what features are driving the individual predictions of these DUC inventory wells. Based on this, we can find opportunities to optimize completions design even further. In this context, we will use the Novi Shapley data to gather information on the model’s predicted contribution of the well’s features to the given production of the well…e.g. how much does proppant contribute to the wells overall prediction relative to the other well data? Similarly, we can evaluate spacing and other model data inputs.

For example, looking at the geologic features such as thickness and depth of the Wolfcamp A-XY we see two relationships emerge:

1. As the thickness of Wolfcamp A-XY increases the relative percentage of increase in 2-Yr cumulative oil volume given the larger completion job begins to increase. This is shown by the delta between the blue dots (larger proppant loading) and the green dots (smaller proppant loading).
2. As the depth of the Wolfcamp A-XY decreases, the expected production from the wells begins to decrease as much as 20% from the model dataset average.

Novi Shapley data also provides valuable insight into the impact of well spacing. As an example, Novi calculates Closest Lateral Distance for every well in the training dataset as the distance in feet from the closest lateral to that well. This was one of the more influential Novi Spacing features in terms of explaining production variance in the training wells.

A couple of things become evident when evaluating the Novi Shapley values for the Closest Lateral Distance datapoint:

1. Unbounded wells, which are vertically orientated in the chart below at 5,000 feet, have an expected uplift in production for the larger completion design that make sense. The variance in well productivity at the larger job size are likely due to subsurface conditions.
2. As the Closest Lateral Distance decreases to 1,500 feet or lower, you can see from the chart below that the Novi model believes that larger completion job has a diminishing uplift as spacing is tighter. It is particularly acute at very tight spacing.
3. Looking at fluid loading ratios in the second chart below, you can see that as Closest Lateral Spacing decreases, there is a very rapid decline in production uplift associated with the fluid intensity ratio relative to proppant. This suggests that increasing fluid ratio relative to proppant will very likely not be economically viable, depending on where in the basin you are.

Risking returns based on model confidence

When creating prediction set outputs in Novi Forecast Engine, one can select two different confidence intervals. We submit that the relative confidence you have in any prediction is extremely important information in evaluating the best solution — engineers can trip themselves (and multi-billion projects) up when they fail to take into account uncertainty. In the example below, we requested that Novi Forecast Engine output the P90 /10 predictions to get an understanding of the model’s confidence, as well as the P35/P65. These choices are configurable with each Prediction Set you create with Novi Forecast Engine.

With each well receiving a prediction at each Confidence Interval (as well as the default, which is P50) we can review the distribution of the results between the bands and determine by the tightness of the projections how confident the model is in the P50 prediction. The tighter the ratio of P90 to P10, the more confident the Novi model is in that prediction. This is shown in the second chart below for a single well case.

We can determine two things from evaluating the P90/P10 confidence ratio:

1. Across the XEC DUC inventory, the model is most confident in the 2,500 lbs/ft job with a 1.2 fluid ratio (first chart below).
2. The model is more confident in predictions for wells in the “Core” of the play, versus predictions made for wells in the fringe (second chart below – lower numbers are better).

When we optimize the P50 output for IP720 oil we see that for most wells the most optimal design is 3,000 lbs/ft job.

However, if we handicap the Novi recommendation based on the relative confidence the model has in each prediction (the P90/P10 ratio), the recommendations change substantially.

To see these outputs in more detail watch the video below where we walk through the analysis and risk adjust completion recommendation for XEC’s DUC inventory.

Conclusions

• Cimarex has ~50 DUC locations spread out across the Delaware basin. They have historically completed these wells with a 2,500 lbs/ft proppant job which aligns with industry but have taken a lower view on fluid ratio (0.8 fluid ratio).
• Novi Shapely data drives important insight in to the relationships between depth, reservoir thickness, spacing and stimulation intensity. Understanding these complex relationships is critical to the DUC completion optimization workflow.
• Novi Shapley data shows empirically that depth and thickness are positively correlated to production; however, this relationship varies based on stimulation intensity.
• The effect of tighter spacing varies based on completion intensity and quality of rock. Completion designs should be tuned to both spacing and subsurface variations across XEC’s DUC inventory.
• Based purely on production uplift, the most optimal completion design on average across XEC’s DUC inventory is the 3k lbs/foot job.
• Model confidence is a proxy for risk. It should be factored in when assigning completion designs across XEC’s DUC inventory.
• Optimal completion designs should take not only engineering and subsurface considerations, but also the relative confidence the model has in its predictions.

Analysis like this can be used to further study and drive better business decisions to maximize returns. If you would like to learn more about our oil and gas well planning software or have questions about this analysis, drop us a line and we will get in touch.

John Chaplin

John Chaplin is a Technical Advisor at Novi where he applies his background in A&D evaluations to ensure Novi's predictive analytics and economics help customers develop better decisions in the field. Prior to Novi, John held various roles in A&D groups at SIPC, Merrill Lynch and Morgan Stanley.