Skip to content Skip to navigation

Machine Learning & Softness: Characterizing local structure and rearrangements in disordered solids

Figure shows an analysis of a polycrystalline material (created via Molecular Dynamic simulations) using ML and the concept of softness.This IRG focuses on the mechanical behavior of disordered materials, particularly beyond the onset of yield. The Figure shows recent advances in using Machine Learning (ML) methods to characterize the local structural environment of disordered materials with respect to susceptibility for particulate rearrangements using a quantity called softness. (A-D) shows an analysis of a polycrystalline material (created via Molecular Dynamic simulations) using ML and the concept of softness [1]. The Figure shows that softness (bright spots in D) is able to capture rearrangements measured as shown by colored particles in (C). This approach correctly identifies crystalline and grain boundary regions as having low values and high variability of softness, respectively. We also extended the concept of softness to anisotropic particles [2] (E). Similar predictive performance to isotropic particles is observed and a recursive feature elimination (RFE) method is introduced to better understand how softness arises from particular structural aspects that can be systematically tuned e.g. by particle aspect ratio.  Indeed, longer particles lead to different global flow patterns for a pillar under compression (F).

 Raw images of the pillar with particles colored by their D2min value, according to the color bar.  Note that the macroscopic response, as measured by pillar shape, is the same for monomers and dimers, but is different for ellipses.

(Figures: A-D: Shows an analysis of a polycrystalline material (created via Molecular Dynamic simulations) using ML and the concept of softness; E&F:Top-down sketches of the particles used in; Bottom panel: Raw images of the pillar with particles colored by their D2min value, according to the color bar.  Note that the macroscopic response, as measured by pillar shape, is the same for monomers and dimers, but is different for ellipses.)

1 - T. A. Sharp, S. L. Thomas, E. D. Cubuk, S. S. Schoenholz, D. J. Srolovitz and A. J. Liu, Machine Learning Determination of Atomic Dynamics at Grain Boundaries, Proceedings of the National Academy of Sciences USA 115, 10943 (2018).  http://dx.doi.org/10.1073/pnas.1807176115
2 - M. Harrington, A. J. Liu and D. J. Durian, Machine Learning Characterization of Structural Defects in Amorphous Packings of Dimers and Ellipses, Physical Review E 99, 022903 (2019).  http://dx.doi.org/10.1103/PhysRevE.99.022903