Relative permeability analysis in the field begins with compiling Special Core Analysis (SCAL) experimental data on core samples. Conventional methods categorize samples by parameters, including rock quality index (RQI), flow zone indicator (FZI), or Winland R35, based on porosity and permeability. Samples are binned by parameter ranges, and collectively analyzed to derive representative permeability curves. The curves receive endpoint analysis, normalization, and denormalization for different rock type bins as per the previously mentioned parameters. However, this relative permeability analysis is a tedious task and requires significant time. Hence, this paper presents a robust and effective machine learning (ML) based approach to derive the relative permeability data sets readily for reservoir engineering study tasks.
This paper presents a Machine Learning (ML) based approach by compiling a database of laboratory-derived SCAL experiments. Thirty-seven experimental oil-to-water relative permeability datasets were collected, which comprised of around 350 data points specific to sandstone reservoir settings. Subsequently, residual oil (Sorw) and irreducible water (Swir) values were tabulated for each core sample. The ML regression models were trained to predict Sorw and Swir using core porosity and permeability as feature variables. Subsequently, core porosity, core permeability, and water saturation (Sw) from relative permeability (kr) experiments were incorporated as features to model krw and kro in the regression models. The trained ML models were then used to further predict the krw and kro curves for any core porosity and permeability for varying water saturation points/steps.
It's often observed that multiple relative permeability curves arise when dealing with varying rock properties, such as permeability and porosity. However, when preparing a bin of rock types, we typically rely on an averaged relative permeability curve for each rock type based on porosity and permeability ratios. This averaging process often necessitates extensive manual calculations and can be quite time-consuming. In this paper, we present an approach that allows for the prediction of two-phase oil and water relative permeability across a range of datasets derived from specific reservoirs with different pore geometries. The derived curves can be effectively utilized in reservoir simulation exercises. We also compare these proposed curves to those generated using the conventional method of averaging relative permeability curves through a modified Brooks-Corey model.