Friction dataset

The datasets contain images of surfaces with structures generated by simplex noise. The initial systems consist of two surfaces of \(\alpha\)-quartz (SiO\(_2\)), where the simplex noise is carved out on the lower surface. The systems are created using molecular-builder . To generate the dataset, we apply Ovito for rendering the images, and transform the images to binary images using Python. The relevant Python scripts can be found at my github repo. All images have shape (1, 128, 128), e.g. one channel.

Labeled dataset
We label the different surfaces by the friction response. When labeling, we measure the slope of the static friction, the maximum static friction and the dynamic friction. This is done using molecular dynamics simulations in LAMMPS. Therefore, each labeled surface is associated with three targets. 

The labeled dataset is named simplex_friction_128x128.pbz2, and contains 1800 surfaces labeled with the friction response. The surfaces were generated on a grid with scales 1-9 and octaves numbers 1-4, 50 surfaces in each category. To unzip and read the dataset in a Python script, use:

import bz2
import _pickle as cPickle

with bz2.BZ2File('simplex_friction_128x128.pbz2', 'rb') as f:
    dataset = cPickle.load(f)

inputs = dataset['x']
labels = dataset['y']

where inputs and labels are Python lists.

Unlabeled dataset
The unlabeled dataset is stored as unlabeled_simplex_friction_128x128.pbz2. It contains 180,000 surfaces generated on a grid with scales 1-9 and octave numbers 1-4, with 5000 surfaces in each category. To unzip and read, use:

import bz2
import _pickle as cPickle

with bz2.BZ2File('unlabeled_simplex_friction_128x128.pbz2', 'rb') as f:
    dataset = cPickle.load(f)

inputs = dataset['x']
attr = dataset['y']

where inputs and attr are Python lists. attr contains information about the simplex noise (scale, octaves, seed).