README.md

---
license: apache-2.0
tasks:
  - materials-simulation
  - energy-prediction
  - force-prediction
  - stress-prediction
  - fine-tuning
frameworks:
  - ase
language:
  - en
tags:
  - OneScience
  - oxide
  - PBE
  - equation-of-state
  - ASE DB
  - materials-science
---

# Oxides

## Dataset Description

Oxides is derived from the oxide polymorph study published by Mehta, Salvador, and Kitchin in 2015. FAIR Chemistry provides the JSON data from the paper's supporting information with its fine-tuning tutorial. This repository extracts `PBE/EOS/calculations` from that data and standardizes it as ASE SQLite databases.

The dataset covers five oxides—`IrO2`, `RuO2`, `SnO2`, `TiO2`, and `VO2`—across 30 oxide/polymorph groups. Each record contains a periodic structure, DFT total energy, atomic forces, and stress. The splits contain 238 training structures, 28 validation structures, and 29 test structures, for a total of 295 equation-of-state (EOS) crystal structures.

Source paper: P. Mehta, P. A. Salvador, and J. R. Kitchin, *Identifying Potential BO2 Oxide Polymorphs for Epitaxial Growth Candidates*, ACS Applied Materials & Interfaces 6(5), 3630–3639 (2015).

Paper: <https://pubs.acs.org/doi/10.1021/am4059149>  
DOI: <https://doi.org/10.1021/am4059149>

Data sources:

- FAIR Chemistry fine-tuning tutorial: <https://github.com/FAIR-Chem/fairchem/blob/main/docs/core/common_tasks/fine-tuning/fine-tuning.md>
- Supporting information distributed by FairChem with the tutorial: <https://github.com/FAIR-Chem/fairchem/blob/main/docs/core/common_tasks/fine-tuning/supporting-information.json>

This repository provides the preprocessed data in ASE DB format. `data/OXIDES/` contains four data files totaling approximately 605 KiB.

## Supported Tasks

This standardized dataset contains oxide PBE EOS training, validation, and test data suitable for:

- Analyzing relationships among oxide EOS structures, energies, and volumes
- Training, fine-tuning, and evaluating interatomic potentials on energies and atomic forces
- Independent validation and generalization analysis of oxide polymorphs
- Validating data loading and processing pipelines for energy, force, and stress labels
- The OneScience eSEN oxide PBE fine-tuning example

## Dataset Format and Structure

The data files are located under `data/OXIDES/`:

```text
data/OXIDES/
├── manifest.json
└── prepared/
    ├── train.db
    ├── val.db
    └── test.db
```

| File | Format | Shape / contents | Description |
| --- | --- | --- | --- |
| `data/OXIDES/prepared/train.db` | ASE SQLite | 238 structures | Training split |
| `data/OXIDES/prepared/val.db` | ASE SQLite | 28 structures | Validation split |
| `data/OXIDES/prepared/test.db` | ASE SQLite | 29 structures | Test split |
| `data/OXIDES/manifest.json` | JSON | Split and source metadata | Data source, random seed, split strategy, and sample counts |
| `metadata/schema.yaml` | YAML | Data schema | ASE DB fields, units, and split definitions |
| `metadata/sha256_manifest.txt` | SHA256 | Four entries | Data-file integrity manifest |

The split is grouped by `oxide/polymorph` with a random seed of `42`. The EOS trajectory for a given polymorph never crosses the training, validation, and test splits, preventing trajectory-level data leakage.

### ASE DB Format

Each row in an ASE SQLite database represents one periodic crystal structure:

| Field | Type | Description |
| --- | --- | --- |
| `numbers` | int array | Atomic numbers |
| `positions` | float array | Cartesian coordinates in Å |
| `cell` | float array | Unit-cell matrix in Å |
| `pbc` | bool array | Periodic boundary conditions |
| `energy` | float | DFT total energy in eV |
| `forces` | float array | Atomic forces with shape `[N, 3]` in eV/Å |
| `stress` | float array | Six-component Voigt stress in eV/ų |
| `oxide` | str | Oxide chemical formula |
| `polymorph` | str | Polymorph group name |
| `xc` | str | Exchange-correlation functional, fixed to `PBE` |

Stress values in the source supporting information are given in GPa and were converted to the ASE unit of `eV/ų` before publication. This repository does not redistribute the original JSON file.

## Usage

Download the dataset:

```bash
hf download OneScience-Group/oxides --repo-type dataset --local-dir ./oxides
cd oxides
```

Validate the dataset directory, sample counts, ASE fields, and SHA256 checksums:

```bash
python scripts/validate_oxides.py \
  --dataset-root data/OXIDES \
  --checksum-manifest metadata/sha256_manifest.txt
```

To skip SHA256 verification and check only the data structure:

```bash
python scripts/validate_oxides.py \
  --dataset-root data/OXIDES \
  --skip-checksum
```

The dataset can be used directly with interatomic-potential training, fine-tuning, and independent evaluation workflows that support ASE DB files. The OneScience eSEN oxide PBE example uses the three databases under `data/OXIDES/prepared/`.

## Official OneScience Resources

| Platform | OneScience main repository | Skills repository |
| --- | --- | --- |
| Gitee | <https://gitee.com/onescience-ai/onescience> | <https://gitee.com/onescience-ai/oneskills> |
| GitHub | <https://github.com/onescience-ai/OneScience> | <https://github.com/onescience-ai/oneskills> |

## Limitations and License

This repository covers only five oxides and a specific collection of PBE EOS trajectories. Its small scale does not represent the general distribution of inorganic materials. It is suitable for tutorials, fine-tuning pipeline checks, and method reproduction, but should not be used directly as a general-purpose interatomic-potential training set.

This release is distributed under the [Apache License 2.0](LICENSE), which applies to the curation, conversion, metadata, validation scripts, and documentation in this repository. The original PBE EOS data comes from the paper's supporting information distributed with the FairChem tutorial. The Apache License 2.0 does not replace the provenance, attribution, or any other applicable requirements of that upstream data. When using or redistributing the dataset, retain the appropriate citations to FairChem, the supporting information, and the paper by Mehta et al.