The data contained in this record is organized in two folders.

1) upload.tar.xz

This compressed folder contains the melting MD simulations of Au nanoparticles containing 147, 309, 561, 923, 2869, and 6266 atoms, carried out in LAMMPS via machine learning force fields.

The simulations are found in the “simulations” subfolder, and are organized first by number of atoms, then by potential employed, finally by simulation. The name of the folder of each simulation contains indications regarding the starting temperature, the final temperature, and the number of simulation steps. Inside each simulation folder there is a .dump file containing the MD trajectory, a .out file containing a summary of temperature, pressure and energy, a .data file containing the starting configuration, and a .in file containing the input script for the simulation.

The folder also contains the three machine learning force fields (lda, rpbe, mixed) we employ to carry out the MD trajectories, both as LAMMPS pair-style tabulated potentials, that can be used by installing the FLARE-LAMMPS plugin, and in the FLARE_compatible .json formats.



2) training_database.tar.xz

This compressed folder contains frames extracted at regular time intervals from an ab initio MD trajectory where an Au NP containing 309 atoms (approx 2 nm in diameter) with an initial FCC morphology undergoes melting from 300 K to 1200 K.
Atomic forces and energy associated with each configuration are calculated within the density functional theory framework, and employing LDA and GGA-rPBE pseudopotentials to generate the training sets for the LDA an rPBE ML-FFs, respectively.
We carry out LDA calculations using the Vienna Ab initio Simulation Package with projector-augmented wave pseudopotentials.
The energy cut-off of the plane-wave basis set was 240 eV, and the tolerance for self-consistency for the electronic steps was set at 10^-6 eV.
We calculate GGA rPBE reference energies and forces using CP2K 6.1. 
All elements are described with the DZVP-MOLOPT basis set with cores represented by the dual-space Goedecker-Teter-Hutter pseudopotentials.
The plane-waves cut-off is set to 500 Ry with a relative cut-off of 50 Ry.
The self-consistent cycle converges when a change of less than 10^-6 eV is observed in the estimate of the system's energy.
The training data is organized in two sub-folders: one for LDA data and the other for rPBE data, and the trajectories are stored in a FLARE-compatible .json files named “flare_structures.json”.