Dataset for "Does fluid structure encode predictions of glassy dynamics?"

This folder contains data in support of "Does fluid structure encode predictions of glassy dynamics?", T. M. Obadiya and D. M. Sussman, arXiv preprint arXiv:2211.00604, (2022). It has three subfolders (described below); all data has been stored in Python's pickle format using protocol version 4, with the pickled files structured as dictionaries.

## "Trajectories/" folder

Each file in this folder is the raw saved output of a molecular dynamics simulation of an 80:20 Kob-Andersen mixture. To generate the trajectories of particles in this mixture, we first equilibrated a system with random initial conditions for 5000 tau at a temperature of T=0.45. We used the final configuration of this as an initial seed for our other simulations: a snapshot was loaded as the initial configuration for our other simulations, each of which was allowed to equilibrate for 1000 tau at its target temperature. The simulations were done in an NVT ensemble with the coupling constant of the thermostat set to 10 tau. The simulations were evolved using a timestep of dt = 0.001 tau, with frames saved every 1 tau.

Each file name contains the temperature at which the simulation was run, and all simulations are of a set of N=4096 particles. The pickle dictionary for each file contains three elements.
"Box_size" contains an array of three elements describing the cubic box simulation box's size along the x, y, and z axes.
"Particle types" has a list of 4096 integers with 0 corresponding to a particle of type A and 1 corresponding to a particle of type B.
"Positions" contains the particle positions for every frame in the saved trajectory.

## "Training_data_V1/" folder

This folder contains the training data for five training temperatures used in the above paper. The paper considered two different training sets (defined by using either phop or cumulative squared displacement as the dynamical label), and the training sets corresponding to this choice of dynamical label are in the corresponding subfolders.

Within each subfolder, the file name indicates the temperature at which the data were obtained. The pickle dictionary for each file contains two elements.
"X_train_rawdata" for every particle in the training set, this contains a list of 100 local structural features (corresponding to the AA and AB local radial distribution functions described in the paper above -- all particles in the training set are particles of type A); the "rawdata" part of the name indicates that these features have not been standardized.
"Y_train" is the corresponding dynamical label, with 1 indicating a large value of dynamical label and a -1 indicating a small value of the dynamical label.

## "Training_data_V2/" folder

This folder contains similar training data for the same five training temperatures, but does not make any assumptions about how to define the local structural features. The pickle dictionaries have the same names as in the Training_data_V1, and the "Y_train" element has the same structure.

Here, though, "X_train_rawdata" for each element contains a list of 4-vectors that correspond to the positions of all particles within 5 sigma of the particle whose dynamical label is being considered. The first 3 elements of each 4-vector give the relative separation between the target particle and its neighboring particle, and the 4th element is the particle type of the neighboring particle (0 for particles of type A, and 1 for particles of type B).