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

AI summary

Overview

• Research area: Applied machine learning for oil and gas production optimization, specifically gas lift optimization in unconventional fields (Bakken).
• Technical level: Intermediate. The abstract describes a two-component system (a forecasting model plus an optimization framework) without mathematical detail, so it is readable by engineers and data scientists but assumes familiarity with artificial lift and optimization concepts.
• Scope: The paper describes an automated, data-driven workflow that predicts gas lift performance from historical production data and uses it to choose optimal gas injection rates, validated in a field pilot and then deployed at scale.

What This Paper Is About

Gas lift wells need the right amount of gas injected to produce optimally, but determining that rate traditionally requires downhole gauges or multi-rate well tests. Those measurements are expensive or physically impractical in many unconventional fields. The paper's goal is an automated workflow that instead learns the relationship between gas injection and production from historical production time series data, then solves for the best injection rates while respecting what the surface facilities can handle.

Key Contributions

1. A machine learning model that forecasts the Gas Lift Performance Curve using only historical production time series — no downhole gauges and no multi-rate well tests are required.
2. A Bayesian Optimization framework that searches for optimal gas injection rates subject to facility capacity constraints.
3. An integrated, automated workflow that combines the forecasting model and the optimizer into a single data-driven process rather than a manual, measurement-dependent one.
4. Field validation and scale-up: the workflow was piloted on 30 wells across 5 well pads in the Bakken, then fully deployed across 200+ gas lift and plunger-assisted gas lift (PAGL) wells.

Main Findings

• Measured production uplift: The pilot on 30 wells across 5 well pads in the Bakken produced an average production uplift of more than 5%.
• No downhole instrumentation needed: The ML model relies on historical production time series data, explicitly avoiding the need for downhole gauges or multi-rate well tests.
• Constraint-aware optimization: The Bayesian Optimization component solves for optimal gas injection rates under facility capacity constraints, meaning the recommendations respect surface infrastructure limits.
• Pilot led to full-scale deployment: Following the pilot's success, the workflow was deployed across 200+ gas lift and PAGL wells in the Bakken.
• Claimed transferability: The authors present the workflow as an effective and economic solution for other assets where downhole data or multi-rate testing are unavailable or infeasible due to cost or facility constraints.
• Not reported in the abstract: The abstract gives no information on model architecture, training data volumes, how uplift varied from well to well, comparisons against alternative optimization methods, or the cost of the workflow itself.

Methodology in Plain English

The approach rests on substituting prediction for measurement:

1. Use data that already exists. Instead of installing gauges or running deliberately varied injection tests, the workflow draws on the historical production time series that wells generate during normal operation.
2. Learn the response curve. A machine learning model is trained to forecast the Gas Lift Performance Curve — essentially how production responds as gas injection rate changes.
3. Optimize within limits. A Bayesian Optimization framework uses that learned curve to search for the best injection rates, while staying inside the constraints imposed by facility capacity.
4. Test, then scale. The authors first ran the workflow as a pilot on a limited set of wells and pads, observed the production response, and then rolled it out across a much larger well population including plunger-assisted gas lift wells.

The abstract does not describe the model's feature set, the optimization objective in detail, or how the workflow is operated day to day.

Why This Matters

Impact on research: The paper demonstrates a pattern of replacing expensive physical measurement with a learned surrogate plus a sample-efficient optimizer. That combination — a data-driven performance curve feeding a Bayesian optimization loop under hard operational constraints — is relevant beyond gas lift to any industrial setting where direct experimentation is costly but operational history is abundant.

Real-world applications:

• Gas lift wells in unconventional fields that lack downhole gauges and cannot justify multi-rate testing.
• Plunger-assisted gas lift (PAGL) wells, which the deployment explicitly includes.
• Assets where multi-rate testing is blocked by cost, surface facility limitations, or logistics rather than by geology.
• Any field with substantial historical production data but sparse downhole instrumentation, which describes a large share of operating unconventional wells.

Industry relevance: The value proposition is economic. Avoiding gauges and well tests removes both capital and deferred-production costs, and automating the rate-selection process reduces the engineering labor needed to manage large well populations. The move from a 30-well pilot to 200+ wells is itself evidence that the operator found the workflow worth institutionalizing, and the authors frame it as applicable to other assets facing the same data and cost constraints.

Future Directions

• Broader deployment beyond the Bakken: The abstract asserts the workflow suits other assets with similar data limitations, but gives no evidence of results outside the Bakken; testing that claim is the obvious next step.
• Robustness as wells change over time: The abstract does not address how the model handles declining production, well interventions, or changing reservoir and facility conditions, or whether it requires retraining.
• Uplift distribution and attribution: Only an average uplift figure is reported. Open questions include how consistent that uplift is across individual wells and pads, and how much of it is attributable to the ML-optimized rates versus the optimization framework or the act of intervening.
• Comparison with established practice: The abstract does not report any comparison against conventional gas lift optimization methods or against multi-rate-test-based workflows, leaving the relative advantage unquantified.

Target Audience

Production and artificial lift engineers working unconventional assets; data scientists and ML practitioners applying optimization to industrial operations; petroleum engineering researchers interested in data-driven workflows; and operations or asset managers evaluating whether to replace gauge-based or test-based gas lift optimization with a data-driven alternative.

Authors’ abstract

In this paper, we present an automated data-driven workflow using Machine Learning (ML) for gas lift optimization in unconventional fields. This workflow integrates a ML model that accurately forecasts the Gas Lift Performance Curve, and a Bayesian Optimization Framework to solve for the optimal gas injection rates under the constraints of facility capacity. The ML model leverages the historical production time series data without requiring downhole gauges or multi-rate well tests. We piloted this workflow on 30 wells across 5 well pads in Bakken and obtained >5% production uplift on average. With the success of the pilot, we have now fully-deployed this workflow in Bakken across 200+ gas lift and plunger-assisted gas lift (PAGL) wells. Moreover, the ML-based gas lift optimization workflow presented in this paper is an effective and economic solution for other assets where downhole data or multi-rate testing are not available/feasible due to cost or facility constraints.