HYBRID FRAMEWORK FOR GAS LIFT OPTIMIZATION: INTEGRATING PHYSICS-BASED MODELS AND DATA-DRIVEN METHODS

Abstract

Gas lift is a widely applied artificial lift method for enhancing oil production in mature and unconventional wells. However, its performance is sensitive to operating conditions, well architecture, and reservoir properties. Predicting production response and identifying optimal injection strategies remain challenging due to complex multiphase flow behavior and variability across wells. Data-driven models often achieve high accuracy within known conditions but struggle to generalize, while physics-based models are computationally intensive and limited in operational coverage. This thesis addresses these challenge by presenting a hybrid framework that integrates data-driven machine learning with physics-based simulations to improve production prediction and support gas lift optimization. The work was structured into two complementary tracks. Track A applied machine learning models, including artificial neural networks, random forest, and XGBoost, to real field data from multiple gas lifted wells. Under conventional 80/20 random splits, all models achieved strong predictive performances, with R2 values approaching 0.97, RMSE of 25 STB/d, and MAE of 16 STB/d. Artificial neural network provides the best overall performance. However, under strict leave-one-well-out (LOWO) evaluation, several wells exhibited negative R2 values and increased prediction errors. The results indicated limited generalization, particularly for wells with distinct or underrepresented operating histories and distributions. Track B served as an automated physics-based modeling workflow using nodal analysis. A framework was developed with the PIPESIMTM Python Toolkit to generate synthetic datasets across a wider operating space. A design of experiments structure was used based on Latin Hypercube Sampling to generate over 6,000 simulation cases, capturing both gas lift and non-gas lift conditions. Additional simulations were conducted across a refined test matrix for wells needing gas lift to construct gas lift performance curves and evaluate the effects of gas injection rate. In general, gas lift shows the greatest benefit in wells with lower reservoir pressure, higher wellhead pressure, heavier oil, and lower solution GOR. The results were then used to identify distinct production response regimes, including monotonic improvement, plateau behavior, and post-peak decline associated with over-injection. The integration of the two tracks through a staged transfer learning approach significantly improved cross-well generalization. While both baseline and transfer learning approaches performed well under standard cross-validation, a clear distinction was observed under LOWO conditions, particularly for wells where baseline model failed. Under LOWO, the hybrid model achieved an average R2 of 0.41, compared to 0.15 for the baseline model, with MAPE decreasing from 56% to 38% and similar reductions in MAE and RMSE. The improvement resulted from expanded operating condition coverage in the hybrid dataset, enabling the model to better represent system behavior across wells.