An initial aproach for particle packing using LSTM networks. The results and methods are described in the paper "Evaluation Of A Particle Packing Method Using Deep Learning", published at the The XL Ibero-Latin-American Congress on Computational Methods in Engineering (CILAMCE 2019).
This project presents a particle packing approach based on the Long Short-Term Memory (LSTM) recurrent neural network architecture. An essential task for the simulation of discontinuous media that use the Discrete Element Method (DEM) is the generation of an initial set of particles, which represent the discontinuous media of interest, with their corresponding positions and radii. The literature presents several packing strategies for different particle geometries such as disks (two-dimensional representation) and spheres (three-dimensional representation). In this context, this paper aims to evaluate the use of deep models based on the LSTM neural network architecture for particle packing. The proposed strategy is comprised of the following steps: a) collecting training data from models by employing any particle packing method, such as the Simple Sequential Inhibition (SSI) or the Poisson Disk Sampling; b) training several variants of LSTM networks to generate the particles by experimenting different combinations of hyper-parameters values; c) generate examples through the proposed algorithm, to evaluate the trained networks. The methodology was implemented using the Keras and the Tensorflow libraries to build, train and evaluate the neural networks. Examples using the different network configurations are presented in order to evaluate the accuracy and applicability of the proposed method