Dataset description

Context and methodology

This dataset was created as part of a study "From Local Structure to Thermodynamics and Transport of Water with Machine Learning Force Fields" evaluating how the exchange-correlation functional used in density functional theory influences the microscopic structure, thermodynamic properties, and transport behavior of liquid water described by machine-learning force fields.

The dataset contains molecular dynamics simulation trajectories and derived analysis data for several machine-learning force fields trained on reference calculations using different density functional theory exchange-correlation (XC) functionals. Results from the machine-learning models are compared with simulations using the classical SPC/E water model.

The investigated XC-functionals include:

• PBE

• PBE-D3

• PBE-TS

• RPBE

• RPBE-D3

• R2SCAN+rVV10

• vdW-DF-cx

The purpose of the dataset is to support the structural, thermodynamic, and transport-property analyses presented in the associated publication. In particular, the data were used to calculate radial and orientational pair-correlation functions, three-body structural descriptors, tetrahedral order parameters, excess-entropy contributions, self-diffusion coefficients, and viscosity.

For each model, five independent molecular dynamics trajectories were generated using different initial velocity seeds. Atomic configurations were sampled every 10 fs, and each deposited trajectory contains 100,000 snapshots.

The original machine-learning force-field simulations were performed using VASP. VASP stores trajectory data in XDATCAR files. To enable analysis with Python, including functionality provided by the MDAnalysis package, the original trajectories were converted to the LAMMPS trajectory format (.lammpstrj).

The full-resolution XDATCAR files contain configurations sampled every 1 fs and are approximately 60 GB per file. Because of their size, these files are not included in the public repository. Instead, the repository contains the converted .lammpstrj trajectories sampled every 10 fs. The original XDATCAR, OUTCAR, and other VASP output files are available from the authors upon request.

Stress-tensor data were recorded every 1 fs and are provided separately from the atomic trajectories.

Dataset structure

The repository is organized by simulation model and by data-processing stage. A representative directory structure is shown below:

.
├── Data/ # NVT raw data for different functionals/force fields. 
│   ├── SPCE/                                                                        
│    │    ├── trajectory/ # Trajectory files from 5 independent simulations.
│    │   │   ├── trajectory_1.lammpstrj
│    │   │   ├── trajectory_2.lammpstrj
│    │   │   ├── ...
│   │   └── stress/ # Stress tensor files from 5 independent simulations.
│    │   │   ├── stress_tensor_SPCE_1.txt
│    │   │   ├── stress_tensor_SPCE_2.txt
│    │   │   ├── ...
│   │   └── PCF/ # Generated PCFs for different nr of configurations/snapshots within independent simulation trajectories. 
│    │   │   ├── trajectory_1/
│    │    │    │    ├── PCF_SPCE_trajectory_1_analyzed_02000.npy
│    │    │    │    ├── PCF_SPCE_trajectory_1_counts_analyzed_02000.npz
│    │    │    │    ├── ...
│    │    │    │    ├── PCF_SPCE_trajectory_1_analyzed_100000.npy
│    │    │    │    ├── PCF_SPCE_trajectory_1_counts_analyzed_100000.npz
│    │   │   ├── trajectory_2/
│    │    │    │    ├── PCF_SPCE_trajectory_2_analyzed_02000.npy
│    │    │    │    ├── PCF_SPCE_trajectory_2_counts_analyzed_02000.npz
│    │    │    │    ├── ...
│    │    │    │    ├── PCF_SPCE_trajectory_2_analyzed_100000.npy
│    │    │    │    ├── PCF_SPCE_trajectory_2_counts_analyzed_100000.npz
│    │   │   ├── .../
│    │    │    │    ├── ...
│   ├── PBE/
│   ├── PBE-D3/
│   ├── PBE-TS/
│   ├── RPBE/
│   ├── RPBE-D3/
│   ├── R2SCAN+rVV10/
│   └── vdW-DF-cx/
├── Analysis/ # Analysis scripts to compute transport properties, three-body structure, pair correlation functions and entropy. 
├── Plotting/ # Final data output and generated figures.
├── Training_data/ # Training data and trained force fields
│   ├── PBE/   
│   ├── ...
├── LICENSE-CODE
├── LICENSE-DATA
└── README.md
└── REQUIREMENTS.txt

The Data/ directory contains the deposited atomic trajectories and stress-tensor data. The directory structure shown for SPCE/ is repeated for the other models. It also contains the cumulative pair correlation functions calculated from the trajectories. Files and folders containing the suffixes _1, _2, _... correspond to the  independent molecular dynamics trajectories. Model names in filenames and directory names identify the corresponding force field. For example, stress_tensor_RPBE-D3_2.txt contains stress-tensor data from the second independent RPBE-D3 trajectory.

The Training_data/ directory contains the ab-initio simulation trajectories ML_AB used for the training of the force fields ML_FF. 

The experimental reference data for the radial distribution functions of water provided in Plotting/Experiment is taken from (CC BY 3.0):
Soper, A. K., The Radial Distribution Functions of Water as Derived from Radiation Total Scattering Experiments: Is There Anything We Can Say for Sure?, International Scholarly Research Notices, 2013, 279463, 67 pages, 2013. https://doi.org/10.1155/2013/279463

File formats

The repository contains the following principal file types and data formats:

• .lammpstrj: Atomic trajectories in LAMMPS trajectory format.

• .txt: Stress-tensor data and other plain-text numerical output.

• .csv: Processed tabular data.

• .npy: Multidimensional NumPy arrays.

• .py: Python analysis scripts.

• .ipynb: Jupyter Notebook analysis scripts.

• ML_AB: Atomic configurations used to train the machine-learned force fields.

• ML_FF: Trained machine-learned force fields.

The .lammpstrj files contain the simulation timestep, simulation-cell dimensions, atom identifiers, atom types, and atomic coordinates for every stored configuration. The stress-tensor files contain the time-dependent tensor components sampled every 1 fs. 

Software requirements

The analysis scripts were written in Python 3.10.12 and use packages including:

• NumPy

• SciPy

• pandas

• Matplotlib

• MDAnalysis

Exact software versions and dependencies are listed in requirements.txt.

The original molecular dynamics simulations using the generated ML-FFs were performed using VASP. Access to VASP is not required to inspect or analyze the deposited .lammpstrj trajectories. Simulations for SPCE were performed using LAMMPS. However, VASP is required to run new molecular dynamics simulations using the trained ML-FFs. The ML_AB files can be inspected with a standard text editor. Details can be found in the VASP Wiki.

Additional resources

The Analysis/ directory contains scripts for:

• calculating the six-dimensional pair correlation functions (PCFs);

• evaluating translational and orientational excess entropy from the generated PCFs; 

• calculating the radial distribution functions;

• calculating three-body angle and tetrahedral-order distributions;

• calculation hydrogen-bond characteristics; and

• calculating self-diffusion coefficients and viscosity.

Further methodological details are provided in the associated scientific publication and its Supporting Information.

The Plotting/ directory contains the script that generates the figures shown in the publication. 

Machine-readable citation information is provided in CITATION.cff.

Reuse considerations

The deposited trajectories are subsampled versions of the original VASP trajectories. Atomic positions are included every 10 fs, whereas the original XDATCAR files contain configurations every 1 fs. The deposited trajectories are suitable for the structural, entropy and transport analyses described in the associated publication. Analyses requiring atomic coordinates at a temporal resolution finer than 10 fs may require the original VASP output files, which are available upon reasonable request. Stress-tensor data are provided at a temporal resolution of 1 fs because this higher sampling frequency is required for the Green–Kubo viscosity analysis.

Each model was simulated using five statistically independent trajectories. The trajectory index should therefore be retained when estimating means, standard deviations, or uncertainties. The independent trajectories should not be treated as a single continuous trajectory. The processed data were derived from the deposited trajectories using the scripts included in the repository. 

Users should consult the associated publication and Supporting Information for complete definitions of the structural descriptors, entropy expressions, convergence procedures, and transport-property calculations.

Licensing and citation

The simulation data, processed data, metadata, figures, and documentation are licensed under the Creative Commons Attribution 4.0 International License. See LICENSE-DATA.

The analysis scripts, conversion scripts, and other source code are licensed under the MIT License. See LICENSE-CODE.

When reusing the dataset or source code in scientific work, please cite the archived dataset and the associated publication using the citation information provided in CITATION.cff and the repository README.