TY  - JOUR
AU  - Nawab, Sohail
AU  - Ali, Muhammad
AU  - Hullio, Imran Ahmed
AU  - Noshad
AU  - Liu, Ning
AU  - Jokhio, Sarfraz A.
PY  - 2026
DA  - 2026/06/19
TI  - A Machine Learning Framework for Artificial Lift Method Selection with Physics-Informed Data Balancing
JO  - Reservoir Science
T2  - Reservoir Science
JF  - Reservoir Science
VL  - 2
IS  - 3
SP  - 228
EP  - 260
DO  - 10.62762/RS.2026.704585
UR  - https://www.icck.org/article/abs/RS.2026.704585
KW  - artificial lift selection
KW  - machine learning
KW  - screening criteria
KW  - class imbalance
KW  - multi-class classification
KW  - ESP
KW  - gas lift
KW  - rod pump
KW  - PCP
KW  - domain-informed data augmentation
AB  - The selection of optimal artificial lift methods using machine learning remains challenging due to complex interactions among reservoir characteristics, fluid properties, and operational constraints. Conventional approaches rely on engineering expertise and static screening criteria, often insufficient to capture multifactorial dependencies. This study presents a framework for classifying the most suitable lift method from four common techniques: ESP, Gas Lift, Rod Pumps, and PCP. A dataset of 990 wells with twelve physically meaningful parameters was compiled, including depth, temperature, GOR, API gravity, reservoir pressure, water cut, production rate, viscosity, sand production, deviation, H\(_2\)S presence, and formation type. To address class imbalance, SMOTE and the proposed DI-SDG method—which derives feature-specific perturbation limits from published intra-class variability data—were evaluated. Six ML models were trained using five-fold stratified cross-validation. Random Forest achieved the best test performance (accuracy: 91.41%, precision: 92.47%, recall: 92.67%, macro-\(F_1\): 0.9255), with XGBoost (\(F_1\): 0.9125) and Gradient Boosting (\(F_1\): 0.9006) also performing well. Generalization was validated via blind well testing using independent field cases. Analysis of 18 misclassified samples showed most errors occurred in overlapping operating envelopes, particularly between ESP and Gas Lift in intermediate GOR ranges. Misclassified samples averaged 64% confidence versus 82% for correct classifications, suggesting a 75% threshold for cases requiring additional evaluation. Overall, the physics-informed balancing approach provides an accurate, interpretable framework for artificial lift selection with reliable field-data performance.
SN  - 3070-2356
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Nawab2026A,
  author = {Sohail Nawab and Muhammad Ali and Imran Ahmed Hullio and Noshad and Ning Liu and Sarfraz A. Jokhio},
  title = {A Machine Learning Framework for Artificial Lift Method Selection with Physics-Informed Data Balancing},
  journal = {Reservoir Science},
  year = {2026},
  volume = {2},
  number = {3},
  pages = {228-260},
  doi = {10.62762/RS.2026.704585},
  url = {https://www.icck.org/article/abs/RS.2026.704585},
  abstract = {The selection of optimal artificial lift methods using machine learning remains challenging due to complex interactions among reservoir characteristics, fluid properties, and operational constraints. Conventional approaches rely on engineering expertise and static screening criteria, often insufficient to capture multifactorial dependencies. This study presents a framework for classifying the most suitable lift method from four common techniques: ESP, Gas Lift, Rod Pumps, and PCP. A dataset of 990 wells with twelve physically meaningful parameters was compiled, including depth, temperature, GOR, API gravity, reservoir pressure, water cut, production rate, viscosity, sand production, deviation, H\(\_2\)S presence, and formation type. To address class imbalance, SMOTE and the proposed DI-SDG method—which derives feature-specific perturbation limits from published intra-class variability data—were evaluated. Six ML models were trained using five-fold stratified cross-validation. Random Forest achieved the best test performance (accuracy: 91.41\%, precision: 92.47\%, recall: 92.67\%, macro-\(F\_1\): 0.9255), with XGBoost (\(F\_1\): 0.9125) and Gradient Boosting (\(F\_1\): 0.9006) also performing well. Generalization was validated via blind well testing using independent field cases. Analysis of 18 misclassified samples showed most errors occurred in overlapping operating envelopes, particularly between ESP and Gas Lift in intermediate GOR ranges. Misclassified samples averaged 64\% confidence versus 82\% for correct classifications, suggesting a 75\% threshold for cases requiring additional evaluation. Overall, the physics-informed balancing approach provides an accurate, interpretable framework for artificial lift selection with reliable field-data performance.},
  keywords = {artificial lift selection, machine learning, screening criteria, class imbalance, multi-class classification, ESP, gas lift, rod pump, PCP, domain-informed data augmentation},
  issn = {3070-2356},
  publisher = {Institute of Central Computation and Knowledge}
}