Improving Machine Learning Predictions of Rock Electric Properties Using 3D Geometric Features

Imaging technology is constantly improving and enabling accurate, deterministic simulations of transport properties through the pore space of the imaged rock sample. Meanwhile, data-driven machine learning has emerged as an alternate tool for modeling transport properties that, once trained, use a fraction of the computational resources that traditional simulations require. However, machine learning models often fail to strictly enforce the physical constraints of the system, leading to solutions that are less accurate than that of traditional solvers. Here we propose a novel hybrid workflow that combines machine learning and conventional simulation methods. The workflow begins with a three-dimensional, binary image of a sample. A trained convolutional neural network extracts spatial relationships between the porous medium geometry and the electrostatic potential field and predicts the electrical properties through a new medium. Instead of assuming a linear potential gradient, this prediction is used as the initial condition of a validated finite difference solver. The implementation of this workflow can improve the simulation run time by an order of magnitude for small images. The success of the proposed workflow heavily depends on the accuracy of model prediction. We previously developed successful methods for prediction of the velocity field (and permeability) of a Newtonian fluid in a porous medium in the laminar regime. Here, we extend the method to predict the electrical potential field. We explore one strategy of improving a model's ability to generalize to unseen samples by supplying geometric characterizations of the pore space. We find that models trained with these features individually do not result in an improvement over the baseline model trained with only the binary image. However, they do provide the model with relational information that can be incorporated into future models. Analysis of electrical properties is one of the most common methods of delineating hydrocarbon saturation in reservoir rock. The proposed workflow helps accelerate the calculation of the electric potential field and can lead to estimating hydrocarbon saturation in real time. We also expect that this workflow is easily generalized to many other transport problems in porous media.