Friction dataset

Friction dataset

This dataset consists of 20,000 two-level topologies labeled by their static friction coefficient. Each image is binary with a resolution of 128x128. The topologies are carved out using simplex noise (see details below) with a fixed porosity of 0.4. The static friction is measured in molecular dynamics simulations. To get started, download the dataset (zip-file) and unzip it:

unzip -d friction_dataset friction_dataset.zip

Structure of dataset

We have chosen a sample oriented folder structure that is easy to inspect. This means that all the information about a sample is collected within its respective subdirectory. We also provide both the binary image, which is both to be used for inspection and automatized reading. The structure looks like the following:

├── extract_dataset.py
├── create_topology.py
├── samples
│ ├── sample_0
│ │ ├── attr.txt
│ │ └── binary.png
│ ├── sample_1
│ ├── ...
│ └── sample_19999
└── README.md

We provide a simple script for extracting the dataset into NumPy arrays in Python (`extract_dataset.py`). From there, the arrays can easily be converted to PyTorch or TensorFlow tensors for machine learning purposes.

Simplex parameters

The topologies are generated by simplex noise. Simplex noise is a procedural noise class known to generate topologies reminiscent of landscapes. We have here used Open Simplex Noise, featuring several parameter: the number of octaves, lacunarity and persistency. Additionally, we have introduced a 'scale' parameter for chaning the size of the topologies (the cluster size). When thresholding the topologies (going from continuous values to binary values), we choose the treshold such that the porosity becomes 0.4 using a simple gradient method. Lastly, the seed may be specified. Simplex noise is deterministic within a specific implementation, meaning that the exact same topology is generated given the number of octaves, scale, lacunarity, persistence, scale and seed. For each samples, these parameters are provided in 'attr.txt', making it possible to regenerate the topology. See `create_topology.py` for more information.

Measuring friction

We measure friction in molecular dynamics simulations. To do this, we carve out the topology from a alpha-quartz (SiO2) surface and slide it across a flat surface of the same material. The static friction is taken as the maximum lateral force (during material failure). For simulation details, see https://arxiv.org/abs/2401.05206.