| Abstract |
Quantifying mineral volumes is crucial for accurate characterization of complex and unconventional reservoirs. Multimineral analysis is a technique used to estimate the fractions of minerals and fluids present in a reservoir based on measurements obtained from well logs. This approach assumes that the log response at each depth can be represented as a linear combination of the individual responses of different logging tools to each constituent, with each component weighted by its relative volume fraction. Achieving accurate mineral quantification necessitates having the number of well logs equal to or greater than the number of minerals in the reservoir. However, in complex reservoirs, there are situations where the number of well logs falls short of the number of minerals, resulting in an underdetermined system of linear equations. In such cases, accurate mineral quantification becomes challenging. To address this challenge, we propose the application of machine learning algorithms to predict mineral volumes using quad combo logs, which include gamma ray, deep resistivity, bulk density, and neutron porosity measurements. Four machine learning algorithms — Artificial Neural Networks (ANN), Random Forest (RF), Adaptive Boosting (AdaBoost), and Support Vector Regressor (SVR) — were utilized to generate predictions for mineral volumes. The performance of each machine learning model was assessed using regression statistical measures. RF delivered the best performance with a correlation coefficient of 0.76. Other algorithms yielded correlation coefficients between 0.11 and 0.54 with the SVR having the lowest value. The results show that this approach offers a promising solution for accurate mineral volume estimation in cases where the number of well logs is insufficient compared to the number of minerals in the reservoir. |
, Ilyas Mellal 
, A. Dehdouh 
