Recurrent neural network model of density relaxation in monodisperse granular systems

• Published in: Computational particle mechanics 2024, Vol.11 (3), p.1119-1132
• Main Authors: Ratnaswamy, V., Rosato, A. D., Chung, Y., Dye, J., Horntrop, D. J., Blackmore, D. L., Ching, N.
• Format: Article
• Language: English
• Subjects: Bulk density; Classical and Continuum Physics; Computational Science and Engineering; Crystal dislocations; Engineering; Initial conditions; Neural networks; Physics; Progressions; Recurrent neural networks; Theoretical and Applied Mechanics
• ISSN: 2196-4378; 2196-4386
• DOI: 10.1007/s40571-023-00676-w

We report on the development of a recurrent neural network (RNN) that models the density relaxation process in initially disordered assemblies of monodisperse spheres within a tapped, three-dimensional container. The RNN model is trained on coordinate data sets generated from granular dynamics simulations to examine microstructure development. In particular, the physics-based model is designed to simulate the evolution of bulk density (characterized by the average solids fraction) within a laterally periodic computational volume starting from an initial, random arrangement of its spheres. Drastically different progressions of individual realizations were observed, often commensurate with the sporadic occurrence of pronounced jumps in density over the duration of a small number of taps. This behavior is consistent with a collective reorganization process previously reported in the literature as an inherent physical feature of the density relaxation process. Visualizations further reveal the formation of crystalline regions separated by dislocations that facilitate bulk sliding motion in the system. To understand how initial conditions and system parameters influence this phenomenon, a considerable amount of data is needed. However, the physics-based simulations necessary to collect this data are too computationally demanding and time consuming. To address this shortcoming and provide a platform for future work, we develop a surrogate RNN model and assess its fidelity with the original physics-based model. Our results suggest that such a surrogate model has the potential to be an important tool in granular systems modeling and research.