Machine Learning Proxy Modeling for Production Forecasting using Ridge-Regularized Polynomial Regression for Deepwater Commingled Reservoirs

Summary Accurate long-term production forecasting is critical for optimizing field development strategies in complex deepwater environments like the Gulf of Mexico/America (GOM), particularly when using commingled production schemes. Traditional numerical simulations are often computationally prohibitive for extensive scenario analysis. With this study, we develop and evaluate machine learning (ML) proxy models as a computationally efficient alternative for predicting 30-year cumulative production. A comprehensive data set is generated using a full factorial design, by varying eight key geological and reservoir parameters across three levels (6,561 simulations). Several ML models are trained and tested, including linear regression, polynomial regression, ridge-regularized polynomial regression, and random forest regression. While random forest yields the lowest initial prediction error (2.6%), diagnostic testing reveals limitations in robustness, particularly for interpolation and extrapolation. The ridge-regularized third-order polynomial model, despite a slightly higher test error (9%), demonstrates superior stability and reliability across the parameter space and beyond training boundaries. Shapley analysis identifies net-to-gross (NTG) ratio and lower zone thickness as the most influential parameters, consistent with prior studies. To demonstrate the practical value of this rigorous algorithm evaluation, the validated ridge-regularized proxy model is then deployed to perform a Monte Carlo simulation. This application proves the selected ML architecture can efficiently and reliably generate thousands of long-term production forecasts to quantify uncertainty, serving as a robust replacement for computationally prohibitive numerical simulators and significantly enhancing decision-making for deepwater field development.