PREreview del A Machine-Learning-Based Gas Lift Optimization Workflow for Unconventional Fields

Summary

The authors present a machine learning workflow for gas lift optimization in unconventional oil fields, applied to the Bakken. The approach combines an ML forecaster that predicts Gas Lift Performance Curves from historical production data with a Bayesian optimization framework that allocates gas injection rates across wells while respecting compressor capacity limits. The method is notable because it does not require downhole gauges or multi-rate well tests, which are often impractical in unconventional assets. A pilot on about 30 wells reported a production uplift above 5%, and the workflow has since been deployed on more than 200 gas lift and PAGL wells. The paper is well organized and tackles a real problem in the field.

Strengths

This work addresses a real operational challenge in unconventional fields, where instrumentation is limited by cost and reliability. Using readily available surface data and historical production trends to forecast gas lift performance is practical and scalable. Combining ML forecasting with constrained optimization under compressor capacity limits gives a complete, deployable solution rather than only a theoretical idea. The pilot results and full-field deployment provide encouraging evidence that the workflow works. The SHAP feature importance analysis helps explain the model. The deployment architecture, with separate training and prediction pipelines, is well thought out and addresses model maintenance.

Major comments

1. Model validation details are insufficient. The paper reports RMSE values of 54.6 bbl/d for the training set and 58.6 bbl/d for the test set (ML Forecaster section). However, typical production rates for these wells are not clearly stated, so it is hard to judge whether this error is acceptable. Please provide average production rates and consider reporting normalized metrics such as MAPE or R-squared. Also, the test RMSE is very close to the training RMSE, which could mean the hold-out period is not very different from the training conditions.

2. A comparison with simpler approaches is missing. The paper notes that physics-based methods are expensive and that data-driven methods typically need downhole data, but no direct comparison is given. Can the authors estimate the production uplift a simpler rule-based approach (for example, keeping a fixed gas-liquid ratio based on past performance) would achieve on the same wells? This would show the added value of the ML and Bayesian optimization components.

3. The Bayesian optimization formulation lacks detail. The paper states that Bayesian optimization is used but does not describe the acquisition function, the prior specification, or how constraint boundaries are handled. If each compressor serves a modest number of wells (for example, 10–30), the dimensionality may be manageable for other optimizers. Please justify why Bayesian optimization was chosen over alternatives such as genetic algorithms or gradient-based methods, and report convergence or efficiency metrics.

4. The production uplift claims need statistical support. The paper reports “>5% production uplift” in the pilot and “~7% production uplift” for the wells optimized after deployment (Results section). However, no confidence intervals, significance tests, or details on how uplift was measured are provided. Oil production varies naturally and follows a declining trend, so distinguishing a 5–7% uplift from noise requires careful analysis. Please clarify whether the comparison is against a forecasted baseline, pre-optimization production, or offset wells, and provide uncertainty bounds on the uplift estimates.

5. The extrapolation ability of KRR is claimed but not demonstrated. Figure 6 conceptually shows tree-based models flattening outside the training range while the KRR stack extrapolates more smoothly. However, no quantitative evidence shows that this improves actual optimization performance. How often are the optimal gas injection rates outside the range seen in the training data? What happens when the model extrapolates too far and suggests a gas rate that is not physically realistic?

Minor comments

• The paper describes removing “the first few months of data on GL” to avoid transient effects (Filter Data). Please state exactly how many months were removed and whether this was data-driven or based on prior knowledge.

• - Forward-filling imputation is used for temporal features (Impute Data), but the amount of missing data is not quantified. For surface pressure and temperature, how much of the time series was imputed? Excessive imputation could introduce bias.

• - The ensemble combines Random Forest, XGBoost and Kernel Ridge Regression through “simple stacking.” Please clarify the meta-learner and how the weights are determined (fixed or learned).

• - Figure 9 shows SHAP feature importance but does not separate the contributions of different temporal lags. It would be useful to see whether the most recent week dominates or whether older lags contribute meaningfully.

• - The workflow uses a weekly time step. Since compressor adjustments can cause multi-day production losses, would a bi-weekly or monthly optimization cycle be more appropriate for some wells? Please discuss the trade-off between optimization frequency and operational disruption.

• - The paper states that 84% of wells are operating in a near-optimal state (Results section). Please define “near-optimal” quantitatively (for example, within a given percentage of the maximum rate on the Gas Lift Performance Curve).

• - How does the workflow handle wells transitioning between natural flow and gas lift, or between gas lift and PAGL? The change in data distribution during these transitions could affect model performance.

• - The deployment section describes automatic model promotion when a new model outperforms the old one on a hold-out set. What happens if the new model is worse? Is there a fallback mechanism or human review?

• - The paper does not discuss computational requirements. How long do training and optimization take for 200+ wells? This would help other operators judge feasibility.

• - Language: “does not worth the risk” should be “is not worth the risk.” In the data cleaning section, please clarify whether “nullify” means setting values to null, removing them, or replacing them with interpolated values.

Recommendation

Major revision. The workflow is practical and the field deployment is impressive, but the paper would benefit from more rigorous validation, clearer methodology, and statistical support for the production uplift claims. Addressing these points would make the contribution stronger and the results more convincing.

Competing interests: None.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.