Integrating Nature-Inspired and Machine Learning Techniques for Estimating Oil Recovery Efficiency of Carbonated Water Injection: Implications for Oil Production and Uncertainty Analysis

Abstract Energy is inarguably the backbone of the socio-economic development of several countries in the Global North and South. While various energy sources exist, hydrocarbons contribute significantly to the world's energy mix (~80%). Global policymakers have also proposed low-carbon sustainable energy carriers (e.g., hydrogen) production from hydrocarbons to support the ever-increasing energy demand, signifying that increased or steady hydrocarbon production volumes are anticipated in the foreseeable future. Carbonated water injection (CWI) has demonstrated remarkable potential in enhancing oil production while sequestering CO2 as a secondary benefit. However, experimental methods for evaluating CWI performance are labor-intensive, time-consuming, and expensive. Incorrect oil recovery estimates can also compromise the techno-economic assessment of CWI. Against this backdrop, k-Nearest Neighbor (kNN), Support Vector Regression (SVR), and Gradient Boosting Regression (GBR) were integrated with nature-inspired optimization techniques: Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Grey Wolf Optimizer (GWO) to estimate the oil recovery efficiency (or factor) of CWI. The developed models were assessed using statistical measures and joint plots. Subsequently, the influence of the different conditions on oil recovery was ascertained with correlation, permutation importance, and Shapley Additive Explanations (SHAP). The outstanding model was then deployed in conjunction with Monte Carlo simulation to evaluate the uncertainties associated with the independent variables on oil recovery potential and provide insights into the practical implications of the integrated modeling approach. While GBR integrated with GWO and PSO (GBR–GWO and GBR–PSO) emerged as the top-performing models (R2 = 0.90, mean absolute error = 3.35%), kNN and SVR showed significant improvements after optimization, with prediction errors reduced by 37–46% and 46–56%, respectively, demonstrating the effectiveness of metaheuristic tuning for models sensitive to hyperparameters. Most importantly, GBR–GWO provided P10, P50, and P90 recovery estimates for 5,000 randomly generated independent variables, as 57%, 72%, and 86%, respectively, without complex experimental settings, showing the developed model's minimal resource dependency and practical utility. Permutation importance and SHAP analysis revealed that injection rate and salinity/oil density were the most and least impactful variables, respectively. SHAP analysis also indicated that high injection rate, porosity, permeability, and pressure enhance recovery, while low temperature, oil viscosity, and salinity favor oil production. These findings align with experimental outcomes and highlight the potential of the integrated modeling approach to complement laboratory and simulation studies, enabling rapid resource-efficient prediction and optimization of CWI recovery efficiency. The technique could also be adapted in adjacent studies to minimize prediction errors and enhance decision-making.