Perpustakaan Digital ITB

Coal Bed Methane (CBM) is an unconventional gas resource predominantly stored in adsorbed form within the coal matrix, whose production is governed by reservoir and adsorption properties. Accurate estimation of the recovery factor is essential for effective field development planning. This study develops machine learning–based proxy models to forecast CBM recovery factor using five parameters: fracture permeability, matrix permeability, fracture porosity, matrix porosity, and Langmuir adsorption constant. A synthetic single-layer CBM reservoir model was constructed and simulated using CMG-GEM, while experimental designs were generated with CMG-CMOST employing Latin Hypercube Sampling. Two proxy modeling approaches—polynomial regression and artificial neural networks (ANN)—were evaluated. The design of experiment consists of 738 total samples, with 590 training and 148 samples had produced both polynomial regression and neural network proxy model successfully. The reduced simple quadratic polynomial regression model achieved an R2 of 0.819 for training and 0.740 for verification, while the ANN multilayer network with one hidden layer of six nodes achieved a higher R2 approaching 1 and RMSE calculation of this model had successfully shown that this model is not categorized as overfit. The results demonstrate that both model shows promising result but ANN shows more superior accuracy, These findings suggest that machine learning–driven proxy models can serve as efficient tools for CBM recovery forecasting, reducing computational costs while maintaining predictive accuracy.