COMPARATIVE ANALYSIS OF MACHINE LEARNING ALGORITHMS FOR POROSITY MODELING

Traditional modeling in hydrogeology and reservoir simulation involves geostatistical, empirical, and physics-based techniques. Geostatistical modeling requires fitting a model to measured data. However, as the geologic environment becomes complex (heterogeneous and anisotropic), selecting the model parameters that effectively capture the spatial structure of the data becomes difficult. Petrophysical/empirical modeling, on the other hand, could be site-specific and may be inaccurate for regions outside the designed experiment. Physics-based modeling entails solving a set of mathematical equations to build forward models. Forward modeling requires several assumptions and a-prior information, which may not always be available, and the solution may suffer non-uniqueness where several models can fit the same dataset. To overcome these challenges, the machine learning has become an integral component of modeling reservoir properties such as porosity, permeability, and density. Machine learning is data-driven and can effectively handle big data with fewer assumptions of model parameters. In this presentation, we compare the results of the two often-used supervised machine learning methods, namely deep learning neural network (DLNN) and random forest (RF) algorithms, for reservoir porosity modeling using synthetic and real datasets. The datasets consist of gamma ray, density, p-impedance, sonic and porosity logs. The results demonstrated the capability of machine learning techniques DLNN and RF to effectively generate a porosity model with low root mean square error in the range of 0.000118 – 0.0143. However, the DLNN required higher computational power and longer running time compared to the RF method.