This blog post is based on the technical paper ” Cross-Basin Analysis of Production Drivers: Insights from Machine Learning Applied to 100,000 Unconventional Wells” presented at URTeC 2025. Download the full paper here.
Most Petroleum engineers would laugh at you if you suggested that you could learn about well production behavior in the Williston from historic production in the Eagleford however that’s exactly what we’re finding after building a “Giga Model” of all the Unconventional oil wells in the Lower 48. (OK maybe not all of them but still most of them)
I’ve spent the last year digging into this problem, analyzing production data from 100,000 horizontal wells spanning seven major U.S. basins. What I found challenges everything we thought we knew about data-driven reservoir analysis. The industry’s obsession with single-basin studies isn’t just limiting our understanding, it might be actively misleading us about what drives production.
When More Data Makes Things Worse
Here’s the uncomfortable truth: traditional machine learning approaches in unconventional oil and gas are fundamentally broken. They’re excellent at finding patterns, but terrible at understanding causation. The result? Models that confidently predict outcomes based on completely wrong assumptions about what actually matters.
Take the Williston Basin, where a naive analysis might suggest older parent wells somehow boost production in nearby child wells. On the surface, this makes no sense. Older parents should be worse for child wells. But the data seems to be telling a different story, a superficial data analysis effort would likely actively hurt the credibility of data-driven reservoir surveillance if these results were taken at face value.
The real explanation reveals the deeper problem actually is in our dataset. In the Williston at least, there is a correlation between parent age and oil production, however the causal reality is somewhat different. Older parent wells in the Williston were completed with much smaller fracture treatments than modern wells. Their reduced size meant they had less impact on nearby reservoir pressure, making them better neighbors for new wells. The age correlation was spurious, completion size was the real driver.
This kind of misattribution isn’t rare. These sorts of relationships are relatively common within Unconventional basins and while we can mitigate the impacts from obvious busts like this, there are more subtle issues that are harder to find and even harder to deal with.
The Permian’s Dirty Secret
The most damning example comes from the Permian Basin, where operators have unknowingly created a massive analytical bias. They’ve consistently developed their best rock at tight spacing and their marginal rock at wider spacing. Economically, this makes perfect sense. Why waste money drilling closely spaced wells in poor-quality reservoirs?
But from an analytical perspective, it’s created a nightmare. Machine learning models can’t separate the effects of completions and geology from the effects of spacing because they’re perfectly correlated. The algorithms assume tight spacing implies good rock quality and big completions, meaning that models see very muted spacing degradation signals.
This effect is present in all the Unconventional basisns. Economic optimization by operators has systematically corrupted our ability to use data analysis to understand fundamental reservoir physics. More data doesn’t fix this problem, it makes it worse by reinforcing the same biased patterns at larger scale.
Breaking Free from Basin Blinders
The solution required thinking differently about scale and scope. Instead of going deeper into individual basins, I went wider across multiple plays. By combining data from the Permian, Eagle Ford, Williston, Denver-Julesburg, Powder River, Uinta, and Appalachian regions, we can start to see much wider trends that better align with geophysical understanding.
For instance, in single basin models we often do see depth as an important feature. This makes sense as it serves as a proxy for pressure and temperature in these larger scale models. However the cross-basin model is also able to see that depth starts to have a deleterious effect after about 13k’. The largere dataset allows the model to learn the impacts of increased thermal maturity and increasing GOR that come with deeper hydrocarbons.
The Spacing Revelation
One of the first things we examined in these cross-basin models was whether more wells alone could improve the models spacing sensitivity. Could we perhaps do away with our Causal inference modeling approach if we threw 100k wells into the wood chipper?
The results however show that the cross-basin model alone doesn’t solve the spacing issues. The data itself requires us to inform the model about the relationship between our features, how spacing is related to completions and geology. The causal model allows us to do that so we need to keep using that as we go forward. Whatever model methodology or dataset you use, you’ll need to tell the model about these underlying relationships within the data.
In this case the cross-basin model does seem to give us some improvements in spacing responses, particularly in the smaller basins that are too small on their own for their own causal models. Uinta in particular shows great spacing sensitivity that it’s learning from all the other basins.
What Geology Actually Tells Us
Basin-wide and smaller models have blinders for geological properties because feature values within a basin or formation just don’t change that much. This leads to smaller models not caring about the impact of key geological properties. The cross basin models don’t have these limitations. For oil wells, porosity and clay content drive most of the performance variation. Higher porosity universally improves production, as expected from basic reservoir engineering principles. Clay content shows more complex behavior, beneficial for gas production but detrimental for oil.
Gas wells follow different rules, with water saturation and clay content emerging as the primary geological controls. These relationships hold true whether you’re looking at the Marcellus, Barnett, or Haynesville formations, suggesting fundamental physical processes that transcend regional variations.
The formation-level analysis revealed surprising patterns. Austin Chalk in the Eagle Ford consistently outperforms other intervals, while Wolfcamp D formations in both the Delaware and Midland consistently underperform. But these formation effects largely disappear when you account for the underlying petrophysical properties, suggesting that formation names are often proxies for fundamental rock characteristics.
The Completion Optimization Sweet Spot
This cross-basin analysis has also revealed some new options for completion optimization landscape. Proppant intensity shows clear diminishing returns around 3,000 pounds per foot across all basins, but with important nuances. Deeper wells can handle higher proppant loadings more effectively, while longer laterals show reduced sensitivity to completion upsizing.
The interaction effects are particularly revealing. In gas wells, high clay content formations respond much better to intensive completions than low clay formations, possibly due to fracture sealing mechanisms that preserve conductivity in clay-rich rocks.
Single-basin studies have consistently underestimated the effects of completion intensity. A Williston-only model suggested 18% production uplift from optimized completions, while the cross-basin model showed 47% uplift for identical wells. The difference highlights how geographical constraints corrupt our understanding of completion effectiveness.
What This Means for Your Operations
These findings fundamentally challenge how we approach unconventional development analysis.
First, spacing impacts can’t be derived from the data directly as we’ve entangled the reservoir responses from spacing with completions and geology. Throwing more wells at the problem won’t resolve it. If we want data driven solutions to spacing optimization we must embrace methods like causal inference that allow us to figure out these relationships from the data, with some guardrails. Bigger than this though, since these relationships are dependent on completions and geology small sample comparisons of analogs is likely leading to significant bias in development strategy optimization. Tools like cross-basin analysis and causal inference can help to resolve these biases.
Second, completion design requires more geological nuance than most operators currently employ. The depth and clay content interactions suggest that one-size-fits-all completion approaches are leaving significant value on the table.
The Path Forward
The industry stands at an analytical crossroads. We can continue refining traditional approaches, adding more features and more sophisticated algorithms to essentially flawed methodologies. Or we can acknowledge that the scope of our analysis fundamentally limits our insights.
Cross-basin thinking isn’t just about bigger datasets, it’s about breaking free from the biases that economic optimization has embedded in our dataset. It’s about rediscovering fundamental reservoir physics that local optimization has obscured.
The 100,000-well analysis provides a roadmap for this transition, but implementation requires changing how we think about data scope and analytical methodology. It requires accepting that more data isn’t always better data, and that the most sophisticated algorithm can’t overcome fundamentally biased inputs.
The question isn’t whether these insights are correct, the physics and the data alignment are too strong to ignore. The question is whether operators will adapt their analytical approaches quickly enough to capture the value before their competitors do.
Dive Deeper Into the Technical Details
This analysis barely scratches the surface of what’s possible when we break free from traditional analytical constraints. The complete 18-page technical study includes detailed methodology frameworks, comprehensive statistical validation, and step-by-step implementation guidelines that you can apply immediately to your own datasets.
Want access to the full research along with our library of resources? [Get Full Access to Our Content Hub →]
Key Takeaways
- Analysis of 100,000 horizontal unconventional wells across seven U.S. basins found single-basin ML models often misattribute correlations as causal drivers.
- Williston Basin data linked older parent wells to higher child-well oil rates, but smaller historic fracture treatments were the actual driver.
- Permian development practices concentrate tight spacing in high-quality rock and wider spacing in marginal rock, biasing spacing-degradation signals in ML models.
- Cross-basin modeling showed depth acts as a pressure-temperature proxy, but becomes deleterious beyond ~13,000 feet due to thermal maturity and rising GOR.
- Cross-basin models improved spacing sensitivity in smaller datasets, with Uinta showing stronger spacing response learned from other basins.
- For oil wells, porosity and clay content drove most performance variation, with higher porosity consistently improving production.
- For gas wells, water saturation and clay content were the primary geological controls across Marcellus, Barnett, and Haynesville.
- Cross-basin results showed diminishing returns for proppant intensity around 3,000 lb/ft, and a Williston-only model’s 18% uplift increased to 47% cross-basin.
