Off-the-Shelf Machine Learning Predicts Deepwater Standpipe Pressure Without Custom Code

Key Findings · At a Glance

• ▸LightGBM, Extra Trees, XGBoost and Random Forest predicted standpipe pressure with test R² between 0.968 and 0.994 on every well.
• ▸The models were trained on 5,921 real-time datapoints from four directional gas wells in the West Delta Deep Marine concession.
• ▸Measured depth was the dominant predictor, averaging about 40% feature importance, followed by pump flow rate at about 16%.
• ▸Hours on bit and downhole revolutions on bit together averaged about 17% importance, capturing time-dependent effects depth alone misses.
• ▸Weight on bit, surface RPM, torque, D-exponent and mechanical specific energy each contributed roughly 1% or less.
• ▸Simple linear benchmarks reached test R² of 0.863 to 0.957, below every ensemble model on every well.
• ▸Well-04, with the fewest features and most intense drilling parameters, gave the lowest test R² for three of four models, though with low RMSEs.

Opportunity

Standpipe pressure is one of the few continuous windows a drilling crew has into conditions at the bottom of the hole. It reflects the hydraulics of the whole circulating system, and a shift in it can point to lost circulation, formation fracturing or changing mud properties. Better hydraulics means faster drilling and less rig downtime, so both operators and drilling contractors have a direct interest in forecasting it.

Earlier machine learning studies reviewed by the authors already report R² above 0.90 for this task. The open question is therefore less about whether it can be done than about who can do it. This paper asks whether the built-in models of a commercial no-code data science platform, Dataiku DSS, can match that accuracy without any programming.

The team drew real-time data from four directional gas wells drilled in deepwater offshore Egypt, targeting a Pliocene channel-levee complex in the West Delta Deep Marine concession. Well-01X was the first to penetrate the untapped central channel; the other three targeted its flanks. The 5,921 datapoints carried between 12 and 24 input features per well. The team ran all fourteen of the platform's in-memory models, cleaned outliers with range rules, standardised the inputs, used five-fold cross-validation and kept the four best performers.

Why Drilling Machine Learning Has Not Yet Delivered

The authors' review identifies a practical bottleneck rather than a scientific one. The most accurate published approaches combine physics-based models with neural networks, and they depend on intricate model integration and programming in Python, MATLAB or R. That confines them to data experts and keeps them out of the hands of operations engineers. Many studies also leave out the software, code or hyperparameters used, which makes them hard to reproduce. One hybrid model cited in the paper degraded as downhole conditions changed and was limited to three-hour intervals unless recalibrated.

Results

All four ensemble models performed strongly on every well. On held-out test data, LightGBM reached R² of 0.994 on Well-01X and XGBoost reached 0.993 on Well-03. Well-04 returned test R² of 0.968 to 0.975, the lowest for three of the four models; Well-02 matched the overall low with XGBoost at 0.968. Test root-mean-square errors ranged from about 23 to 48 psi, and training R² reached 0.999 in places. The authors conclude that "readily implementable, ensemble models are highly effective for this task".

The Practical Applications

Assessment

The authors acknowledge signs of potential overfitting, with training R² as high as 0.999. The study covers a single field, and they are explicit that "local retraining on new data would be required for other fields", with performance contingent on local data quality. They recommend further outlier removal to tidy some feature effects, and they excluded geomechanical variables because rock properties are hard to measure or predict reliably.

Readers should apply further caution. Each model was trained and tested on data from the same well, so the study does not show prediction on an unseen well or in live, forward-looking operation. Drilling data are sequential, and if the splits were random, neighbouring samples could flatter scores. Measured depth carries much of the predictive weight, so part of the skill may reflect tracking the depth trend. The paper also describes weight on bit, RPM and torque as below 1% importance, although its own Table 17 puts them slightly above 1%. For technology buyers, the defensible reading is that a no-code workflow is a credible starting point; a blind-well test is the obvious next proof point.