AI drilling optimization research from Russia’s Skolkovo Institute of Science and Technology (Skoltech) has produced a striking number. For one of three oil wells studied, a machine-learning model estimated that better drilling settings could raise the rate of penetration, the speed at which the bit cuts down through rock, by up to 68.4%. The same settings were predicted to cut torque on the drill string by 41.2%.
The study, by Assistant Professor Shadfar Davoodi of the Skoltech AI Center, was announced on 7 October 2026 and published in the open-access journal Results in Engineering. This AI drilling optimization work combines three techniques that are usually used separately: prediction, explanation and optimisation across several goals at once. The headline needs care, though. The 68% is a projection from a model, not a result measured at a rig, and the author says so plainly.
This article explains what the AI drilling optimization study did, what it found well by well, and what “up to 68%” really means in hours saved. We also set out the limits and what the method’s design offers businesses outside oil and gas. For a wider view of AI in heavy industry, see our report on AI for oil and gas plant modelling.
Table of contents
- What the AI Drilling Optimization Study Found
- How AI Drilling Optimization Works in This Study
- The Three Settings Behind AI Drilling Optimization
- AI Drilling Optimization Results, Well by Well
- How Accurate Is the Model Behind the AI Drilling Optimization?
- What “Up to 68%” Means in Hours Saved
- The Limits of This AI Drilling Optimization Research
- How AI Drilling Optimization Differs From Earlier Methods
- Why AI Drilling Optimization Matters Beyond Oil
- What Comes Next for AI Drilling Optimization
- AI Drilling Optimization: Frequently Asked Questions
- References
What the AI Drilling Optimization Study Found
The core finding of the AI drilling optimization study is that there is no single best way to drill. The study found that the right combination of drilling settings depends on the rock and on how the well was already being drilled. For two wells the model saw large gains available. For the third it saw almost none, because that well was already being drilled close to its best settings.
| Item | Detail |
|---|---|
| Paper | Drilling performance enhancement in heterogeneous carbonate reservoirs: An integrated machine-learning and multi-objective optimization approach |
| Author | Shadfar Davoodi (Skoltech; Tomsk Polytechnic University), sole author |
| Journal | Results in Engineering, volume 32, article 113178 (open access, CC BY-NC 4.0) |
| Data | 5,738 records from three vertical wells, Asmari Formation, southwest Iran |
| Best model | Long short-term memory (LSTM) network |
| Headline projection | +68.4% rate of penetration, -41.2% torque, -17.9% specific energy in one well |
| Status | Model estimates only; not yet tested in the field |
Where the data came from
Skoltech’s press release says only that the data came from “three vertical wells drilled in a carbonate formation”. The paper’s abstract is more specific. It places the wells in the Asmari Formation of southwest Iran, a limestone unit that hosts many of that country’s largest oilfields. The Asmari Formation is known for rock properties that change sharply with depth, which is exactly the problem the study sets out to solve.
Why carbonate rock is hard to drill well
Carbonate reservoirs are “heterogeneous”, meaning their hardness and structure vary from one metre to the next. Settings that drill one layer efficiently may waste energy or damage the bit in the next. The press release puts the challenge simply: “not simply to drill faster but to find a balance between speed, load, and energy efficiency”.
How AI Drilling Optimization Works in This Study
The AI drilling optimization framework runs in three stages. First, models learn to predict how the well will respond to different settings. Second, explainable AI shows which inputs drive those predictions. Third, an optimisation algorithm searches for the settings that give the best trade-off between several goals.
“We were able to integrate prediction, interpretation, and optimization into a single framework,” Davoodi said in the Skoltech announcement. “The next step is to validate these recommendations under field conditions.”
Stage one: predicting penetration rate and torque
The models were trained to predict two outputs: the rate of penetration (ROP), and torque, the twisting force on the drill string. Inputs combined live drilling parameters with rock properties from a “mechanical earth model”, a description of rock strength and stress along the well.
Before training, the data went through sensitivity analysis, normalisation and recursive feature elimination, which drops inputs that add little. Three model types were built and tuned using particle swarm optimisation: a long short-term memory (LSTM) network, a random forest and a multilayer perceptron, a basic neural network.
Stage two: explaining what drives the model
The study used two explainable AI methods, SHAP and partial dependence plots (PDP). SHAP, introduced by Scott Lundberg and Su-In Lee in 2017, shares out each prediction among the inputs, showing how much each one pushed the answer up or down. The analysis found that rotary speed, measured in revolutions per minute (RPM), was the most influential input in this dataset.
Stage three: searching for the best trade-off
For the AI drilling optimization step, the prediction models were then plugged into NSGA-III, a genetic algorithm for problems with several competing goals, published by Kalyanmoy Deb and Himanshu Jain in 2014. It searched combinations of three controllable settings to raise ROP while lowering torque and mechanical specific energy at the same time.
The Three Settings Behind AI Drilling Optimization
A driller cannot change the rock. What they can change are the settings at the surface. The AI drilling optimization in this paper adjusts three of them, and each one pulls in more than one direction.
| Setting | What it is | Why more is not always better |
|---|---|---|
| Weight on bit (WOB) | The downward force pressing the bit into the rock | Too much raises torque, wears the bit and can stall it |
| Rotary speed (RPM) | How fast the drill string turns | The most influential input here; higher speed can waste energy or cause vibration in hard rock |
| Flow rate | How fast drilling fluid is pumped down to clear cuttings | Higher flow cleans the hole but costs pump energy and pressure |
Mechanical specific energy explained
The third goal, mechanical specific energy (MSE), is the energy spent to remove a unit volume of rock. The idea dates back to R. Teale’s 1965 paper on specific energy in rock drilling. In AI drilling optimization terms, a low MSE means the rig is breaking rock efficiently; a high MSE means energy is going into heat, vibration and wear instead.
The physical floor in the model
The framework also included what the press release calls “a physical constraint reflecting the minimum energy required to break the rock”. That matters. A purely data-driven optimiser can recommend settings that look good on paper but break the laws of physics. The constraint stops the AI drilling optimization from proposing impossible answers.
AI Drilling Optimization Results, Well by Well
The study reports AI drilling optimization results separately for each well, and the differences are the most useful part of the paper. The wells are distinguished by their unconfined compressive strength (UCS), a measure of how hard the rock is.
| Well | Rock strength (UCS) | Projected ROP change | Projected torque change | Other notes |
|---|---|---|---|---|
| Low-strength well | 56 MPa | +68.4% | -41.2% | Specific energy -17.9% |
| High-strength well | 68 MPa | +66.9% | -59.6% | Achieved through a substantial RPM adjustment |
| Moderate-strength well | 58 MPa | +2.7% | Not stated | Already near optimal; RPM reduced; aim shifts to cutting load and energy |
Bar widths are each figure’s size divided by 68.4. All six numbers come from the paper’s abstract and Skoltech’s press release.
Why the moderate well barely moved
The 2.7% result is not a failure. According to the press release, that well “was initially drilled in a regime close to optimal”. In that case the AI drilling optimization recommended slowing the rotary speed and focused on “reducing load and energy consumption” rather than speed. A tool that says “you are already doing well here” is more credible than one that promises 68% everywhere.
No universal recipe
The paper’s own summary is that its AI drilling optimization “revealed formation-dependent drilling strategies”. In plain terms, the softer rock and the harder rock needed different changes. The high-strength well got most of its gain from “substantial RPM adjustment”. This is why the author stresses that the recommendations “depend on rock properties and the initial drilling conditions”.
How Accurate Is the Model Behind the AI Drilling Optimization?
AI drilling optimization projections are only as good as the model that makes them. The paper reports how well the LSTM predicted the two outputs, both on the training wells and on a third well held back entirely for testing.
| Output | R² (training wells) | Typical error (RMSE) | R² (unseen test well) |
|---|---|---|---|
| Rate of penetration | 0.846 | 1.93 metres per hour | 0.805 |
| Torque | 0.995 | 178.15 lbf·ft | 0.977 |
Reading the R² numbers
R² measures how much of the variation in the real data a model explains, where 1.0 is perfect. Torque was predicted almost perfectly. Penetration rate was harder: 0.846 on the training wells and 0.805 on the unseen well. Roughly a fifth of the variation in drilling speed on the new well was not explained by the model.
Why that gap matters for the 68% figure
The optimised settings were judged by the same models. So the 68.4% is a model’s prediction of how a change it recommends would perform, in the part of the data where the model is least accurate. That does not make it wrong. It does mean the AI drilling optimization headline carries more uncertainty than a single decimal place suggests.
What "Up to 68%" Means in Hours Saved
A 68% faster drilling speed from AI drilling optimization does not mean 68% less time. Time is distance divided by speed, so the saving is always smaller than the speed increase. This is the most common way readers overestimate claims of this kind.
The arithmetic is simple. If speed rises by a fraction g, drilling time falls to 1/(1+g) of what it was. A 68.4% rise gives 1/1.684, or 59.4% of the original time: a 40.6% saving. The high-strength well’s 66.9% gives 40.1%. The moderate well’s 2.7% gives 2.6%.
Bar widths are each percentage divided by 68.4.
A worked example
To make that concrete, take a purely illustrative 500-metre section drilled at 10 metres per hour. These are round numbers chosen for the sum, not figures from the paper. At the original speed it takes 50 hours on bottom. At a speed 68.4% higher, 16.84 metres per hour, it takes 29.7 hours. The saving is 20.3 hours, which is 40.6% of the original time.
| Scenario (500 m section, illustrative) | Speed | Hours on bottom | Hours saved |
|---|---|---|---|
| Baseline | 10.00 m/h | 50.0 | – |
| +68.4% (low-strength well) | 16.84 m/h | 29.7 | 20.3 |
| +66.9% (high-strength well) | 16.69 m/h | 30.0 | 20.0 |
| +2.7% (moderate well) | 10.27 m/h | 48.7 | 1.3 |
On-bottom time is not total rig time
There is a second discount. Penetration rate only affects the hours when the bit is actually cutting rock. A well also involves pulling and running pipe, fitting casing, cementing and waiting on equipment, none of which a faster ROP changes. The paper does not estimate total well time or cost, and this article does not either.
The Limits of This AI Drilling Optimization Research
The author is careful about what this AI drilling optimization study shows, and the press release repeats the caveats. They are worth stating plainly, because headlines tend to drop them.
Model estimates, not field results
“The author stresses that these values represent estimates produced by surrogate models rather than the results of full-scale field trials,” Skoltech said. The abstract is just as direct: the figures “have not been validated through field trials”. No rig has yet drilled with the recommended settings to confirm the gains.
One formation, three wells
The AI drilling optimization data are historical, from three wells in one formation. Applying the approach “to other geological settings will require further tuning and validation”, the press release says. Reusing a model trained on one field in another is a known machine-learning problem. Techniques such as transfer learning can help, but they still need new data from the new site.
A small dataset by AI standards
At 5,738 records, the dataset is small next to the volumes behind most modern AI systems. That is normal for drilling research, where good labelled data is scarce. It does mean the results should be read as a promising method rather than a proven tool.
| What the headline suggests | What the paper actually says |
|---|---|
| AI makes drilling 68% faster | A model projects up to 68.4% faster ROP in one of three wells |
| A tested result | Not validated in field trials |
| Works on any well | Recommendations depend on rock and starting conditions; one well gained 2.7% |
| 68% less time | At most about 40.6% less on-bottom time, by our arithmetic |
How AI Drilling Optimization Differs From Earlier Methods
Drillers have tried to optimise penetration rate for more than half a century. What is new in this AI drilling optimization study is not the goal but the combination of methods. Setting it against two well-known earlier approaches shows what has changed.
Regression models from the 1970s
In 1974, Adam Bourgoyne and Farrile Young published a multiple regression approach to optimal drilling in the Society of Petroleum Engineers Journal. Their model links penetration rate to factors such as depth, weight on bit and rotary speed through a fixed equation, with coefficients fitted to field data. It is still taught today. Its weakness is that the shape of the equation is chosen in advance, so it cannot learn interactions it was not designed to capture.
Specific energy surveillance in the 2000s
In 2005, Fred Dupriest and William Koederitz of ExxonMobil described real-time surveillance of mechanical specific energy to maximise drill rates. Instead of predicting ROP, crews watch MSE live and change settings when it rises, a sign that energy is being wasted. The approach is practical and widely used, but it relies on people reading the signal and choosing what to change.
What the new study adds
The Skoltech framework learns the relationships from data rather than assuming them, explains them with SHAP, and lets an algorithm weigh three goals at once. It still uses MSE, the 2005 idea, as one of its targets. In that sense the AI drilling optimization in this paper builds on earlier methods rather than replacing them.
| Approach | How it chooses settings | Main limitation |
|---|---|---|
| Regression model (Bourgoyne and Young, 1974) | Fits a fixed equation for ROP to field data | The equation’s shape is assumed, not learned |
| MSE surveillance (Dupriest and Koederitz, 2005) | Crews watch energy per unit of rock and adjust when it rises | Depends on people interpreting a live signal |
| ML with multi-objective search (Davoodi, 2026) | Learns ROP and torque from data, explains them, searches trade-offs with NSGA-III | Projected only; trained on three wells in one formation |
Why AI Drilling Optimization Matters Beyond Oil
Drilling is not unique. Many industrial processes balance speed against wear, energy and safety. The design of this study, predict then explain then optimise against several goals, carries over well to other fields.
The same pattern in manufacturing
Machining is a close cousin. A cutting tool’s feed rate and spindle speed trade output against tool wear and power use, much as weight on bit and RPM do in a well. The same three-stage approach could find better settings on a production line. Our machine learning model development team builds this kind of model for clients.
Explainability makes adoption possible
In any AI drilling optimization system, the SHAP stage matters as much as the optimiser. A drilling engineer will not change settings on a live well because a black box says so. Being able to show that rotary speed drives the prediction, and in which direction, is what turns a model into a tool people trust. The same is true for any predictive analytics project inside a business.
Constraints keep the answers physical
Building physics into the optimiser is the other lesson. Industrial AI that ignores known limits will eventually recommend something unsafe. For another heavy-industry firm moving from automation to AI, see our report on Caterpillar’s approach to AI deployment.
What Comes Next for AI Drilling Optimization
Skoltech says the approach “is expected to be tested on field data and further developed toward adaptive real-time drilling optimization”. That is the real prize: a system that adjusts settings while drilling, as the rock changes, rather than recommending fixed settings after the fact.
From offline study to real-time control
Real-time AI drilling optimization raises new questions. Live sensor data is noisy, rigs have safety interlocks, and a recommendation that arrives too late is useless. The LSTM chosen here is designed for sequences of data over time, which suits that goal. But getting from a paper to a control room is usually the hardest step.
Questions a buyer should ask
Any business offered an AI drilling optimization product, or any AI optimiser for an industrial process, should ask a few basic questions. Was the result measured in the field or projected by a model? How did it perform on data it had never seen? What happens when conditions fall outside its training data? And what physical limits are built in? This study answers the last three honestly, and is open about the first.
AI Drilling Optimization: Frequently Asked Questions
Did AI make drilling 68% faster?
No. A model projected that better settings could raise the rate of penetration by up to 68.4% in one of three wells. The settings have not yet been tested at a rig.
Who carried out the study?
Shadfar Davoodi, an assistant professor at the Skoltech AI Center in Moscow, with an affiliation to Tomsk Polytechnic University. The paper is published in Results in Engineering.
Which AI methods were used?
An LSTM neural network to predict penetration rate and torque, SHAP and partial dependence plots to explain the predictions, and the NSGA-III algorithm to search for the best trade-off between speed, torque and energy.
Which drilling settings does it change?
Weight on bit, rotary speed (RPM) and drilling fluid flow rate. Rotary speed was the most influential input in this dataset.
Does the method work on any oil well?
Not yet. It was built and tested on historical data from three wells in the Asmari Formation in southwest Iran. Other rock types would need further tuning and validation.
References
Drilling performance enhancement in heterogeneous carbonate reservoirs (Results in Engineering)
Skoltech press release (Scoop)
The concept of specific energy in rock drilling, R. Teale, 1965
An evolutionary many-objective optimization algorithm (NSGA-III), Deb and Jain, 2014
A unified approach to interpreting model predictions (SHAP), Lundberg and Lee, 2017
Rate of penetration (Wikipedia)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
