For decades, artificial lift optimization has relied heavily on manual intervention, periodic well tests, and the judgment of experienced engineers. Electric submersible pumps (ESPs), gas lift, and beam pumps (the three most widely used lift systems) each have presented their own optimization challenges, such as balancing set points, managing compressor constraints, and interpreting dynamometer cards. These processes, while effective, have historically been time-intensive, reactive, and difficult to scale across large well inventories.
The accelerating deployment of machine learning (ML) and automation is changing this landscape. By embedding intelligence into the control loop, operators now can move from reactive decision-making to proactive, continuous optimization. What was once a manual task handled by individual engineers is increasingly performed by autonomous systems that learn from data, adjust in real time, and scale effortlessly across hundreds or thousands of wells.
Recent case studies illustrate how this transformation is unfolding across different lift types, each demonstrating that the age of autonomy in artificial lift is no longer aspirational but is now an operational reality.
Neural-network models trained on ESP data sets in the Permian Basin now recommend and directly write optimal pump set points, as shown in paper SPE 219528. The deployment delivered a 2–4% oil uplift and longer run life, while demonstrating full self-pumping capability, where pumps operate with minimal human oversight.
In gas lift, ExxonMobil’s large-scale rollout of a closed-loop optimization workflow has been equally impactful. As detailed in paper SPE 219553, more than 1,300 wells are now managed by an automated system that runs multirate tests, updates downhole pressure models, and applies ML for wells without gauges. The outcome: a consistent approximately 2% production uplift, achieved with little incremental operational or capital expenditure.
On the beam-pump side, manual interpretation of dynamometer cards is being replaced by real-time ML classification. In paper SPE 224979, more than 700,000 dynacards were analyzed with models such as XGBoost and convolutional neural networks, achieving approximately 95% classification accuracy. By integrating with OSIsoft AF and exception-based surveillance, the solution provides predictive analytics and proactive alerts, reducing downtime and extending equipment life.
When comparing the three approaches, some clear distinctions emerge. In ESP optimization (SPE 219528), operators relied on neural networks to recommend and autonomously adjust set points, achieving production uplift across more than 200 wells. Gas lift optimization (SPE 219553) took a broader fieldwide approach, scaling a closed-loop workflow to more than 1,300 wells, where multirate testing and ML models continuously updated injection rates, delivering uplift at minimal incremental cost. Meanwhile, beam-pump optimization (SPE 224979) focused less on direct production uplift and more on reliability and predictive surveillance, where ML models enabled proactive maintenance and reduced downtime.
Despite their differences, the three initiatives share commonalities: Each system moved from manual, engineer-driven adjustments to automated and autonomous optimization; each leveraged ML in combination with existing field infrastructure; and each demonstrated that even modest per-well improvements, when scaled across hundreds or thousands of wells, translate into substantial fieldwide gains.
Across ESPs, gas lift, and beam pumps, these case studies demonstrate a unifying shift: ML is redefining artificial lift as a self-optimizing, continuously learning system. Production uplifts of 2–4%, when multiplied across fields, translate into substantial incremental barrels. Equipment longevity improves as systems learn to operate within optimal ranges, while engineers are freed from repetitive monitoring to focus on higher-value analysis.
The advances described here are more than isolated technical improvements; they are signposts of a broader digital transformation across upstream operations. Artificial lift, long the domain of manual adjustments and engineer intuition, is now becoming a proving ground for autonomy.
By embedding AI into the very core of production systems, operators are not just optimizing individual wells; they are building the foundation of the digital oilfield of the future, where decisions are data-driven, workflows are autonomous, and engineers focus their expertise on complex challenges rather than routine tuning.
Summarized Papers in This October 2025 Issue
SPE 219528 Machine Learning Optimizes Autonomous ESP Use in the Permian by Ryan D. Erickson, Vital Energy, et al.
SPE 219553 Gas Lift Optimization in the Permian Uses Machine Learning, Artificial Intelligence by Pooya Movahed, ExxonMobil, et al.
SPE 224979 Real-Time Machine Learning Enhances Dynacard Surveillance, Predictive Analytics by Shifa Khaliah Al Kiyumi, Daleel Petroleum, et al.
Fahd Saghir, SPE, is director of artificial intelligence at Occidental Oman. He holds a PhD degree in petroleum engineering from the University of Adelaide and a BS degree in electrical engineering from the University of Houston. Saghir has 19 years of experience in diverse digital roles across the energy, utilities, and natural-resources sectors, with the past 8 years focused on the Internet of Things, applied ML, and data analytics. In his previous role with a national oil company, Saghir worked as a data scientist and digital-transformation enabler, collaborating with engineers, data scientists, and technology specialists. He is an active member of SPE, having served as chairman of the Integrated Digital Solutions Subcommittee (2020–22) and as a committee member for various SPE forums.
