Deep-learning methods for predicting permeability from 2D/3D binary-segmented images

Computational Rock Physics has become a reliable method for obtaining macro-scale properties of rocks from micro-scale physical processes by using computations such as finite element, finite difference, and Lattice Boltzmann methods. Instead of using these conventional numerical simulations, I developed machine learning methods and showed that it is possible to predict 3-D transport properties, using geometrical features from both 2-D and 3-D µXCT binary segmented images. In this work, both multilayer neural network (MNN) and convolutional neural network (CNN) algorithms are employed to predict permeability. Training is performed through both feed-forward and back-propagation with Bayesian Regularization using gradient descent algorithm. The inputs for MNN can be geometrical parameters such as Minkowski Functionals (porosity, specific surface area, integral of mean curvature (for 3-D), and Euler number). For CNN, the inputs are either 2-D or 3-D binary images. Presentation Date: Tuesday, October 18, 2016 Start Time: 11:35:00 AM Location: 143/149 Presentation Type: ORAL