Clastic Reservoir Rock Grain Size Estimation from Wireline Logs Using a Random Forest Model: Initial Results

Summary Grain size is a key input to various reservoir models. The models require a continuous log of grain size. Core samples are usually not available over the entire reservoir section. The most accurate grain size measurement is obtained from sieve and laser particle size analyses. These methods are expensive. The conventional method, the visual core description, is time-consuming, subjective, and nonreproducible. Alternative methods include the use of empirical equations, nuclear magnetic resonance (NMR) relaxation time, and acoustic velocities. These latter methods require inputs that are not sufficiently available, not applicable to different geological settings, or not available for all wells. This paper proposes a new methodology that estimates reservoir rock grain size for a new well or reservoir section from archival core description data and their corresponding wireline logs using machine learning technology. Nine wells from a clastic reservoir are used. Seven wells are combined to build the training set while the remaining two are used for model validation. Three machine learning methods are implemented and trained with optimized parameters. The results showed that, despite the subjectivity and bias associated with the core description data, the machine learning methods are capable of estimating the grain size for the validation wells.