---
license: cc-by-4.0
language:
- en
pretty_name: MLIP-liquid-corrections
size_categories:
- 1M<n<10M
tags:
- chemistry
- molecular-dynamics
- interatomic-potentials
- bsse
- dispersion
configs:
- config_name: atm
  data_files:
  - split: full
    path: "ATM/atm_dataset.csv"
- config_name: bsse
  data_files:
  - split: full
    path: "BSSE/bsse_dataset.csv"
---

# MLIP-liquid-corrections

## Dataset Summary

MLIP-liquid-corrections provides energy and force labels for basis-set
superposition error (BSSE) and Axilrod–Teller–Muto (ATM) corrections. The
molecular configurations are derived from
[OMol25](https://huggingface.co/facebook/OMol25), including SPICE-origin
structures.

BSSE labels were calculated using the counterpoise (CP) method at the
ωB97M-V/def2-TZVPD level of theory. ATM labels were calculated from the
three-body term of D4 using the ωB97M (`wb97m`) parameterization.

The dataset contains 2,026,607 configurations covering H, C, N, O, F, P, S,
Cl, and Br. Energy and force labels describe corrections rather than total
molecular energies and forces.

## Subsets

| Subset | Configurations | Description |
| --- | ---: | --- |
| ATM | 1,606,854 | Additive ATM energy and force corrections. |
| BSSE | 419,753 | Native BSSE energy and force quantities. |

Both subsets use the Hugging Face split `full`. Source IDs beginning with
`train_` do not define a train/test split. Cross-subset records are not aligned
by row number.

## Dataset Structure

Each subset contains a CSV index and one HDF5 file:

```text
ATM/
  atm_dataset.csv
  omol25_spice_cleaned_corrected_atm.h5
BSSE/
  bsse_dataset.csv
  omol25_spice_cleaned_corrected_bsse.h5
```

### CSV Index

Both CSV indices contain exactly these five columns:

```text
h5_key,num_atoms,spice_name,smiles,mapped_smiles
```

`h5_key` is the original source record ID, such as `train_1000`, and names the
corresponding HDF5 group. Select the HDF5 file from the index's ATM or BSSE
subset. `num_atoms` is the configuration's atom count. `spice_name` is the
source SPICE molecular name and may be shared by multiple configurations.
`smiles` and `mapped_smiles` follow the conventions described below.

### HDF5 Format

Both files use `grouped_correction_with_species_v1`, with one top-level group
per configuration. `N` is the number of atoms in that configuration.

**ATM**

```text
/train_1000/
  atomic_numbers  (N,)     uint8
  positions       (N, 3)   float64
  ATM_energy      scalar  float64
  ATM_forces      (N, 3)   float64
  smiles          scalar  UTF-8 string
  mapped_smiles   scalar  UTF-8 string
```

**BSSE**

```text
/train_10000403/
  atomic_numbers  (N,)     uint8
  positions       (N, 3)   float64
  BSSE_energy     scalar  float64
  BSSE_forces     (N, 3)   float64
  smiles          scalar  UTF-8 string
  mapped_smiles   scalar  UTF-8 string
```

Each HDF5 group retains only the fixed-length UTF-8 attributes `smiles` and
`mapped_smiles`. Atom map `j+1` in `mapped_smiles` corresponds to atom row `j`.
The SMILES values in the CSV indices match the corresponding HDF5 group
attributes.

See [DATA_PROVENANCE.md](DATA_PROVENANCE.md) for source and processing details,
[dataset_info.json](dataset_info.json) for statistics, and
[DATA_MANIFEST.sha256](DATA_MANIFEST.sha256) for file hashes.

### Units

| Quantity | ATM | BSSE, native |
| --- | --- | --- |
| Position | Å | Å |
| Energy | eV | Hartree |
| Force | eV/Å | Hartree/Bohr |

For additive BSSE corrections, negate and convert the native labels:

```text
E_additive[eV]   = -BSSE_energy * 27.211386245988
F_additive[eV/Å] = -BSSE_forces * 27.211386245988 / 0.529177210903
```

## License and Citation

[Creative Commons Attribution 4.0 International (CC BY 4.0)](LICENSE).
Preserve [NOTICE](NOTICE) and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md)
when redistributing. See [CITATION.md](CITATION.md) for the OMol25 and SPICE
publications.

When using the BSSE and ATM data, please cite the associated study:

```bibtex
@misc{wang2026subtleenergycorrectionsreshape,
  title={Subtle Energy Corrections Reshape Condensed-Phase Property Predictions from a Machine-Learning Interatomic Potential},
  author={Haihui Wang and Xinyan Wang and Jiashu Liang and Yuanheng Wang and Chang Han and Wen Yan and Tianze Zheng and Kuang Yu},
  year={2026},
  eprint={2611.06109},
  archivePrefix={arXiv},
  primaryClass={physics.chem-ph},
  url={https://arxiv.org/abs/2611.06109},
}
```