Data for "Assessing generative modeling approaches for free energy estimates in condensed matter"

The dataset contains raw molecular dynamics (MD) simulation data used to train the models in the paper "Assessing Generative Modeling Approaches for Free Energy Estimates in Condensed Matter" (arXiv:2512.23930). It includes:

Atomic coordinates

Forces

Potential energies

The data covers monatomic ice in cubic and hexagonal phases, as well as Lennard-Jones solids in HCP and FCC crystal structures. Each combination of structure and system size has two datasets (indicated by the suffixes _001 and _002), which can be used for training and evaluation. Each dataset contains 100,000 samples generated from MD simulations. Frames are saved every 1000 MD steps with a timestep of 1 fs.

Additionally, the dataset includes Jupyter notebooks demonstrating how to:

Load the data

Convert between absolute coordinates and displacements

Compute the free energy of the reference system (Einstein crystal)

Extract the unit cells of the different structures

Run the MD simulations to generate the data