2020/10/02 by Renat Sergazinov, Sergazinov, Renat, Miroslav Kramár +1
Environmental Science · Engineering · #Landslides and related hazards #Rock Mechanics and Modeling #Soil and Unsaturated Flow
paper · pdf · doi:10.48550/arxiv.2010.01163
Photoelastic techniques have a long tradition in both qualitative and\nquantitative analysis of the stresses in granular materials. Over the last two\ndecades, computational methods for reconstructing forces between particles from\ntheir photoelastic response have been developed by many different experimental\nteams. Unfortunately, all of these methods are computationally expensive. This\nlimits their use for processing extensive data sets that capture the time\nevolution of granular ensembles consisting of a large number of particles. In\nthis paper, we present a novel approach to this problem which leverages the\npower of convolutional neural networks to recognize complex spatial patterns.\nThe main drawback of using neural networks is that training them usually\nrequires a large labeled data set which is hard to obtain experimentally. We\nshow that this problem can be successfully circumvented by pretraining the\nnetworks on a large synthetic data set and then fine-tuning them on much\nsmaller experimental data sets. Due to our current lack of experimental data,\nwe demonstrate the potential of our method by changing the size of the\nconsidered particles which alters the exhibited photoelastic patterns more than\ntypical experimental errors.\n