This directory contains two datasets for single-task benchmarks that require additional sampling and splitting preprocessing. Other single-task benchmark datasets have standalone train/val/test splits and are omitted here.

### Liquid Water and Ice Dynamics (`water_ice`)

1. Original Data:

lw_pimd: liquid water at P = 1 bar and T = 300 K, at the PI-AIMD level;
ice_pimd: ice Ih at P = 1 bar and T = 273 K, at the PI-AIMD level;
ice_triple_I: ice Ih at P = 1 bar and T = 330 K, at the classical AIMD level;
ice_triple_II: ice Ih at P = 2.13 kbar and T = 238 K.

These data were used for the two papers:
1) L. Zhang, J. Han, H. Wang, R. Car, W. E, Physical review letters 120 (14), 143001
2) H.-Y. Ko, L. Zhang, B. Santra, H. Wang, W. E, R. A. DiStasio Jr, R. Car Molecular Physics 117 (22), 3269-3281

Please consider citing them when you use the data for your work.

2. Downsample Data:
We uniformly sampled <0.1% of the data (133 structures) for training, with 50 frames allocated for validation and all remaining data reserved for testing, see `random_fold/split.py` for details.

### The Formate Decomposition on Cu (`fcu`)

1. Original Data:

1) Batzner, S. et al. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic
potentials. Nat. communications 13, 2453 (2022).

This dataset consists of configurations characterizing the decomposition process of formate on Cu(110), focusing on C-H bond cleavage.
It includes initial states (monodentate/bidentate formate), intermediate configurations, and final states (H ad-atom with gas-phase CO2).
The Nudged Elastic Band (NEB) method generated reaction pathways, followed by 12 \textit{ab initio} molecular dynamics (AIMD) simulations using the CP2K code.
These simulations produced 6,855 DFT structures with a 0.5 fs time step over 500-step trajectories, capturing dynamic evolution across reaction coordinates.
The dataset provides atomistic-scale insights into catalytic decomposition mechanisms through systematically sampled configurations.

2. Data Split:

The full dataset was partitioned into training (2,500 structures), validation (250 structures), and test (remaining 4,105 structures) sets via uniform random sampling, see `data/downsample/split.py` for details.

