This project aims to use machine learning methods to optimize packing densities in granular matter. Employing dimensional reduction, regression techniques, and numerical optimization, we search for novel dense packing shapes, which are then tested in the Molecular Dynamics simulations using LAMMPS. This project refers to the paper titled "Machine learning approaches for the optimization of packing densities in granular matter", which can be found in https://pubs.rsc.org/en/content/articlelanding/2023/sm/d2sm01430k.
In this project, we use the dataset generated by Roth & Jaeger in https://pubs.rsc.org/en/content/articlelanding/2016/sm/c5sm02335a#! for particle shapes consisting of n = 5 overlapping spheres.
This part provides the codes for the following steps:
- dimensionality reduction of the shape space with principal component analysis
- regression in the reduced shape space
- numerical optimization
to identify novel dense packing shapes.
We provide molecule files and input scripts to run simulations in LAMMPS for the predicted particle shapes obtained in the first part. A molecule file requires to specify the volume and moment of inertia of the molecule (particle consisting of n=5 overlapping spheres). Therefore, we also provide the Monte-Carlo simulation codes (written in Jupiter Notebook) to measure the volume and moment of inertia of the molecules.
We introduce two different methods to calculate packing density: Voronoi and Centroid method. The codes for these calculations are written in Jupyter Notebook.
More details can be found in the Documentation.pdf.
Here, we provide the predicted shapes, their volume and moment of inertia calculated in part 2, the simulation output files for one molecule, and the packing densities for all the molecules measured with the different methods. The Mathematica notebook used to create some of the plots is also included.