Considerations on the ability of Supervised Neural Networks to estimate the 3D shape of particles from 2D projections

Abstract

Equipment to observe and classify granular media is widely used to capture granular systems morphologies. Although some equipment models are very efficient in analyzing samples with numerous grains, these tools present a limitation related to the inability to obtain directly the particle’s 3D shape. To overcome this drawback, different techniques have been applied to estimate the three-dimensional shape of grains from the analysis of their 2D projections. Despite the progress made, because it is an ill-posed problem, the developed methods have not yet provided a definitive solution. In this work, we investigated the ability of Supervised Neural Networks (SNN) to estimate the three-dimensional shape of some particles of Angular sand and Ottawa sand using data from single projections. We performed simulations employing the Discrete Element Method (DEM) and comparisons of shape descriptors to sets of reconstructed 3D particle shapes. The SNNs can adequately correlate and generate new particles from the analysis of 2D sand projections, showing potential applicability in the geometry reconstruction of granular materials.