Two Knobs, One Record: What Shale Design Models Cannot Separate

Chevron says it has a revenue interest in one of every five wells in the Permian, which gives the company visibility into operations and data across the basin, and that connecting artificial intelligence to that presence allows improved well design and execution, reduced costs and enhanced decision-making [1]. The system doing the design work has a name. APOLO, formally the Automated Production Outlook and Location Optimization, is described by Chevron as using AI to analyse relationships among numerous subsurface variables among thousands of wells, and, in Chevron's own words, it shows how changes in spacing, proppant and fluid use may affect production [2]. It runs in the Permian and DJ basins, and Chevron expects to extend it across its global shale and tight assets [2].

Three of Chevron's peers spent 2024 buying development histories. ExxonMobil closed its acquisition of Pioneer Natural Resources on 3 May 2024, taking the combined company to more than 1.4 million net acres in the Delaware and Midland basins with an estimated 16 billion barrels of oil equivalent of resource [3]. Occidental closed CrownRock on 1 August 2024 [4]. ConocoPhillips closed Marathon Oil on 22 November 2024 [5]. Inside seven months, three independents' complete development records changed owner. None of the three closing announcements says anything about what happens to that data afterwards [3][4][5]. ExxonMobil's public account of its digital work describes artificial intelligence and machine learning applied to Permian operations rather than to a well-by-well development history [6], which is a different object. Chevron is the one company here that says plainly what its system reads. So the question is not whether anyone will fit a model to a record like this, because Chevron says APOLO already does. It is what such a record can support when they do.

Read that sequence as a data problem and it looks like the best one in the industry. The U.S. Energy Information Administration counted 17,450 producing wells in the Delaware Basin and 27,540 in the Midland Basin as of June 2021, all installed since January 2011 [7]. Close to 45 thousand wells, every one of them logged, completed and produced under a design someone chose on purpose. Nobody in subsurface machine learning has a corpus like it.

The trouble is what those wells recorded.

Two knobs, one record

The Permian did not vary its design one factor at a time. It ran a decade-long programme in which several things moved together. The EIA's own summary of Permian drilling and completion practice puts the share of wells with 11,000-foot or longer laterals at less than one percent in 2014 and around 20% in 2021, notes that proppant and fluid use has not changed significantly since 2017, and describes multi-well pad drilling as increasing reservoir penetration while posing challenges for well spacing optimisation and production interference during simultaneous hydraulic fracturing. In the first half of 2021, 25% of Midland and Delaware pads carried at least nine wells [7]. Longer laterals, heavier completions through 2017, and more wells packed onto a pad are not independent programme decisions. They arrived together because they were justified together.

Chevron's own sentence about APOLO is the tell. The three design variables it names, spacing, proppant and fluid use, sit in one clause because they sit in one decision [2]. A development plan sets them jointly. The record therefore contains their joint history and not much else.

That is a statistical problem with a name, and it is not the one most subsurface teams are trained to look for.

Why spacing is not just another column

A reservoir engineer will object that spacing is in the model. Chevron says APOLO analyses relationships among numerous subsurface variables [2], and spacing is plainly one of them. The objection is right and it does not close the question, because of what spacing does physically.

A University of Texas at Austin study by Manchanda, Bhardwaj, Hwang and Sharma used a fully coupled three-dimensional poroelastic model to work out why child wells underperform. Reservoir rocks form a geomechanical system that responds to both pressure and stress changes, so long-term production from a hydraulically fractured parent well reduces pore pressure around it and changes the stresses, which alters the pressures and stresses around any child well drilled later. The study names three factors that induce interference between parent and child wells: reservoir depletion, total stress change and effective stress change. The direction is not universal. Parent well production typically benefits from fracturing a nearby child well in the Bakken and Haynesville plays, and typically suffers in the Woodford, Eagle Ford and Niobrara [8].

So a child well's productivity is a function of the state its neighbour left behind. Spacing is not a design variable whose effect adds cleanly alongside proppant loading. It is a variable that moves productivity through the rock, and that moved through the record in step with the completion knob. Those are exactly the two conditions that bias a fitted coefficient, and the bias has a closed form.

The algebra, on one line

Write a well's productivity as a completion effect, a spacing effect and everything else.

$$y_i = \beta_c\,c_i \;-\; \delta\,s_i \;+\; u_i$$

FormulaLaTeX

Copy LaTeX

Here $c_i$ is completion intensity, $s_i$ is how tightly the well was spaced, $\beta_c$ is the true completion effect, and $\delta$ is the parent-child depletion penalty per unit of tightening. Now fit productivity on completion intensity alone. The least-squares slope is not $\beta_c$. It is $\beta_c$ contaminated by the omitted term, weighted by how strongly spacing tracked completion intensity in the record.

$$\hat\beta_{\text{naive}} \;=\; \beta_c \;-\; \delta\,\frac{\operatorname{Cov}(s,c)}{\operatorname{Var}(c)}$$

FormulaLaTeX

Copy LaTeX

Put both variables on a common scale and that ratio is just the historical correlation $r$ between spacing and completion intensity, so the estimate is a straight line in $r$ and it has a root.

$$\hat\beta_{\text{naive}} = \beta_c - \delta r \qquad\Longrightarrow\qquad r^\star = \frac{\beta_c}{\delta}$$

FormulaLaTeX

Copy LaTeX

Nothing about that requires anyone to run a literal one-variable regression. It is the clean limit of a general condition: a fit recovers the completion effect only from variation in completion intensity that spacing did not already explain. Where the two moved together, the model is being asked to attribute a shared movement, and the answer it returns is whatever the historical pairing implies.

The bench

The exhibit below builds a seeded cloud of synthetic wells from that model, plots productivity against completion intensity with each well coloured by which third of the spacing distribution it fell into, and solves both fits live from the plotted sample. The sample is constructed so its correlation equals the control exactly, which makes both fits exact rather than approximate, so the two lines say something about the estimator and nothing about sampling luck.

correlation in the record, r: 0.80 (0 to 0.95)parent-child depletion penalty: 0.85 (0 to 0.9)true completion effect: +0.55 (0 to 0.8, held positive)

The bench opens with the true completion effect held at plus 0.55, a depletion penalty of 0.85, and a record whose spacing and completion intensity correlate at 0.80. The naive fit reads minus 0.13. The spacing-adjusted fit reads plus 0.55, the number that was put in.

Drag the correlation down towards zero and the white line rotates up through horizontal until it lies on the aqua one. Drag it to 0.95 and the white line falls to minus 0.26. Across that entire sweep the aqua line does not move, because it is estimating the coefficient you set rather than the record you drilled. Two of the three controls behave that way: the correlation and the depletion penalty both move the naive line and leave the adjusted line exactly where it was. The third, the true completion effect, moves both, because it is the quantity both fits are trying to recover.

The number worth reading off is the crossing. At the opening settings it is 0.65, printed in closed form rather than searched. A development programme whose spacing and completion decisions tracked each other at 0.65 is not an exotic case. It is one operator, one play, one five-year plan where the same engineering argument justified both knobs.

The honest caveat, which cuts the other way

Adding spacing to the fit does recover the coefficient here, and that matters. The exhibit is not an argument that the problem is unfixable. It is an argument that the fix depends on something nobody prints on a slide: how much variation in completion intensity survives after spacing is accounted for.

That residual is what the two-variable solve divides by. Its determinant is $1 - r^2$, which is 0.0975 at the bench's maximum correlation of 0.95. The variance inflation factor is the reciprocal of that determinant, so at $r = 0.80$ the variance of the completion coefficient is 2.78 times what an orthogonal design would have given and its standard error, being the square root, is 1.67 times. At $r = 0.95$ the variance is 10.26 times and the standard error 3.20 times. The bench's residual is exactly orthogonal by construction, so its adjusted estimate is exact and the inflation costs it nothing. A real fit on a real programme pays that cost in full, on top of everything else.

Which means the two failure modes sit next to each other. Leave spacing out and you get a confidently wrong sign. Put it in and you get the right sign with an interval wide enough to contain most of the decisions the model was bought to make. Neither is a modelling error. Both are properties of a record that moved two knobs at once.

Our own record is the opposite regime, and it is worth saying so

We have written before about evaluating subsurface models honestly when the corpus is small. On a 14-well borehole-image engagement in a Middle East carbonate field, the sharpest question in the whole peer review was whether our train and test splits leaked, because depth-adjacent patches from the same well are near-duplicates and a shuffle can put an almost identical neighbour on both sides of the boundary. Holding out a whole well costs one fourteenth of the corpus, and every well in that field carried its own structural character, so a one-well test set reports the luck of the draw as much as the skill of the model [9].

We spent our evaluation effort on split policy because at 14 wells leakage was the failure that could actually bite. That was the right allocation for that corpus and it is the wrong allocation for this one.

At 45 thousand wells, leakage is not the interesting failure. The record is observational, the two levers were pulled together, and no split policy in the world fixes a confound. Worse, the split hides it. If spacing and completion intensity are correlated at 0.80 in your training wells, they are correlated at 0.80 in your held-out wells too, because both came out of the same development plan. The held-out score will confirm the biased coefficient and report it as generalisation. That is the specific reason a large validation set feels like reassurance here and is not.

What to ask before the design model sets the plan

The useful question for an operator evaluating a system of this kind is not how accurate the forecast is. Forecasting is the part these systems are good at, and Chevron's account of APOLO is careful to describe it as a forecasting and standardisation tool that delivers explainable forecasts and is expected in future to recommend optimal production strategies [2]. The gap between forecasting a well and prescribing its design is the whole subject of this piece.

Three questions close it. What is the correlation between spacing and completion intensity in the wells the model was fitted on, by operator and by vintage? What variance inflation does that correlation impose on the completion coefficient the model is being asked to act on? And where in the acreage did somebody vary completion intensity at fixed spacing, or spacing at fixed completion intensity, on purpose?

The last one is the only question with a clean answer, and it is a drilling decision rather than a modelling one. Chevron says APOLO learns over time and can simulate different designs in different spots [2]. Simulation over a confounded record reproduces the confound at higher resolution. A small number of pads drilled deliberately off the historical correlation would do more for the completion coefficient than another ten thousand wells drilled on it.

Key takeaways

1. 1Chevron holds a revenue interest in one of every five Permian wells and runs APOLO, formally the Automated Production Outlook and Location Optimization, across thousands of wells to show how changes in spacing, proppant and fluid use may affect production. Spacing and completion sit in one sentence because they sat in one decision.
2. 2ExxonMobil closed Pioneer on 3 May 2024, Occidental closed CrownRock on 1 August 2024 and ConocoPhillips closed Marathon Oil on 22 November 2024, so within seven months three independents' complete development histories changed owner. None of the three closing announcements says what becomes of that data, so the question worth asking is what a record of that kind can support rather than who is fitting a model to it.
3. 3Spacing is not an ordinary column. A University of Texas at Austin poroelastic study names reservoir depletion, total stress change and effective stress change as the three factors driving parent-child interference, and finds the sign of the effect on the parent differs between the Bakken and Haynesville on one side and the Woodford, Eagle Ford and Niobrara on the other.
4. 4Omitted-variable bias in this setting is a straight line in the historical correlation, so the fitted completion coefficient crosses zero at r equal to the true effect divided by the depletion penalty. At the exhibit's opening settings that is 0.65, which is well inside what a single operator's five-year plan can produce.
5. 5Putting spacing into the fit recovers the coefficient, at a price the exhibit measures: the determinant is one minus r squared, so the standard error is inflated 1.67 times at a correlation of 0.80 and 3.20 times at 0.95.
6. 6EarthScan's own 14-well borehole-image work spent its evaluation effort on split policy because leakage was the failure that could bite at that size. At 45 thousand wells the confound is the failure, and a held-out well drilled under the same plan carries the same correlation, so the holdout confirms the bias instead of exposing it.

References

[1] Chevron. Built on legacy, driven by discipline: Chevron's Permian advantage explained. 14 August 2025. https://www.chevron.com/newsroom/2025/q3/built-on-legacy-driven-by-discipline-chevrons-permian-advantage-explained

[2] Chevron. How APOLO helps Chevron pinpoint prime drilling locations. 12 November 2025. https://www.chevron.com/newsroom/2025/q4/how-apolo-helps-chevron-pinpoint-prime-drilling-locations

[3] ExxonMobil. ExxonMobil completes acquisition of Pioneer Natural Resources. 3 May 2024. https://corporate.exxonmobil.com/news/news-releases/2024/0503_exxonmobil-completes-acquisition-of-pioneer-natural-resources

[4] Occidental. Occidental Completes Acquisition of CrownRock. 1 August 2024. https://www.oxy.com/news/news-releases/occidental-completes-acquisition-of-crownrock/

[5] ConocoPhillips. ConocoPhillips completes acquisition of Marathon Oil Corporation. 22 November 2024. https://www.conocophillips.com/news-media/story/conocophillips-completes-acquisition-of-marathon-oil-corporation/

[6] ExxonMobil. Digital technologies. https://corporate.exxonmobil.com/who-we-are/technology-and-collaborations/digital-technologies

[7] U.S. Energy Information Administration. Drilling and completion improvements support Permian Basin hydrocarbon production. Today in Energy, 19 October 2021. https://www.eia.gov/todayinenergy/detail.php?id=50016

[8] Manchanda, R., Bhardwaj, P., Hwang, J. and Sharma, M. M. Study Assesses Interference Between Parent, Child Wells In Horizontal Pad Developments. The American Oil & Gas Reporter, Frac Facts, July 2018. https://www.aogr.com/magazine/frac-facts/study-assesses-interference-between-parent-child-wells

[9] EarthScan. Patch Splits or Well Splits? Evaluating Subsurface AI Honestly When You Only Have 14 Wells. https://earthscan.io/insights/patch-splits-or-well-splits-evaluating-subsurface-ai-honestly-with-14-wells