Predicting Sand Production in JAKS Oilfield Using Machine Learning and Geological-Operational Data

Abstract

Sand production is a major operational challenge in weakly consolidated reservoirs where conventional empirical models do not account for the complex nonlinear interactions. A predictive machine-learning framework is developed for JAKS Oilfield to proactively manage sanding and optimize field development. Four algorithms (ANN, SVR, RF and XGBoost) were trained on eight geological and operational features, including depth, porosity, permeability, UCS, pressure depletion and water cut. The results show that XGBoost is better than all others with R2 of 0.951, RMSE of 2.15 kg/day and a classification accuracy of 96.3%. SHAP analysis identifies UCS, depletion and permeability as dominant predictors and establishes a critical UCS threshold of 1200 psi for proactive intervention. With 10% input noise, the XGBoost framework shows good generalization, providing a tool to optimize drawdown and completion design. The findings conclude that this approach significantly surpasses traditional correlations, offering a robust decision-support system for efficient hydrocarbon recovery.