Towards a more accurate characterization of granular media 2.0: Involving AI in the process

Abstract

We introduce a Convolutional Neural Network (CNN) to reduce grains’ manual inspection time after image processing on raw 3D x-ray computed tomography (3DXRCT) images from a sample of granular material to obtain level-set function-based digital twins of individual grains. The CNN automatically distinguishes properly segmented digital grains with up to 90% of accuracy. This algorithm is trained using, ground-truth, level set-based digital grain representations from a natural soil sampled at Jaramijó (Ecuador). The implemented convolutional neural network provides groundbreaking processing power, reducing the, otherwise, manual inspection time expended for a small sample, e.g., 200 000 grains, from approximately a couple of weeks to only a few hours. Furthermore, transfer learning and training from scratch are compared for artificially graded granular materials such as Øysand (Norway) and Hostun sand (France). The CNN’s learning process is interpreted by means of grain morphological parameters, i.e., sphericity, roundness, grain diameter, and volume-surface ratio. Hence, being able to automatically segment a greater amount of grains from 3DXRCT images of natural and artificial soils in a short period of time, enables us, for first time, to perform actual 3DLS-DEM-based virtual laboratory testing (a plug-and-play one-stop shop). Providing unprecedented and unique data for engineering applications.