Title: Cornerstones are the Key Stones: Studying Granular Clogging with Machine Learning as an Experimental Guide

Author (Talk): Sam Dillavou, University of Pennsylvania

Abstract:

The spontaneous arrest of flow through an outlet is a unique feature of granular materials. This process, clogging, has resisted mechanistic understanding because it is extremely difficult to interrogate. It is a collective event whose dynamics depend on a wide range of length- and time-scales as well as poorly understood interactions like friction between grains. I will discuss our progress in overcoming these difficulties using interpretable machine learning (ML) as an experimental guide. We amass a large experimental dataset of (over 50,000) clogging events from a quasi-2D, tri-disperse granular system, and show that a wide range of ML algorithms from Support Vector Machines (SVMs) to deep Convolutional Neural Networks (CNNs) predict these clogs only modestly above chance. However despite this "failure," the solutions found by SVMs still provide meaningful information about the dynamics of clogging; they clearly highlight the importance of grains adjacent to the outlet, that is, potential cornerstones for static arches. We experimentally verify this finding using a single fixed grain, whose position, even when it does not obscure the outlet, can modify flow rate, time, and ejected mass by over a factor of 2.