README.md

Citation & License

Commercial use is strictly prohibited.

If you use this dataset in your research, please cite all the following works:

@article{cao2025beyond,
  title={Beyond Structure: Invariant Crystal Property Prediction with Pseudo-Particle Ray Diffraction},
  author={Cao, Bin and Liu, Yang and Zhang, Longhan and Wu, Yifan and Li, Zhixun and Luo, Yuyu and Cheng, Hong and Ren, Yang and Zhang, Tong-Yi},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2026}
}

Task Overview

This benchmark includes 6 primary tasks for evaluating CPP models in crystal property prediction, divided into regression and classification missions.

Mission Settings

Regression Tasks

1. T1: Formation Energy Prediction (float) | eV/atom
2. T2: Band Gap Prediction (float) | eV
3. T3: Bulk Modulus Prediction (float) | Gpa
4. T4: Shear Modulus Prediction (float) | Gpa
5. T5: Young’s Modulus Prediction (float) | Gpa

Classification Task

7. T6: Metal/Non-metal Classification (int, binary classification)

---

Dataset Description

The datasets are hosted on Hugging Face and stored in ASE database (.db) format. Each database includes crystal structures and corresponding property labels.

fe_bandg/

• Train: MP_100_bgfe_train.db (86,071 crystals)

• Val: MP_100_bgfe_val.db (12,295 crystals)

• Test: MP_100_bgfe_test.db (24,593 crystals)

• Tasks: T1, T2

  • Keys: formation_energy, band_gap

modulus/

• Train: MP_modulus_train.db (6,631 crystals)

• Val: MP_modulus_val.db (947 crystals)

• Test: MP_modulus_test.db (1,895 crystals)

• Tasks: T3–T5

  • Keys: bulk_modulus, shear_modulus, youngs_modulus, poissons_ratio,

metal_nometal/

• Train: MP_100_metal_train.db (86,071 crystals)

• Val: MP_100_metal_val.db (12,295 crystals)

• Test: MP_100_metal_test.db (24,593 crystals)

• Task: T6 (label: 0 or 1)

  • Keys: metal

---

Data Format

All data is stored in ASE database format (.db). Each entry contains both the crystal structure and the associated target properties.

In preparing the dataset for model training, we applied a log10 transformation to the elastic modulus values. This step was chosen because the modulus spans several orders of magnitude, and taking the logarithm helps reduce the scale differences, making the distribution more balanced and easier for the model to learn from. Since the logarithm is only defined for positive values, we excluded any records where the elastic modulus was zero or negative. This filtering step ensured that all remaining data points were valid for the transformation and avoided numerical issues during training. As a result, the processed dataset provides a cleaner and more consistent basis for the machine learning workflow.

---

Example: Reading Data from an ASE Database

You can read data using the ase.db module as follows:

from ase.db import connect

# Load the database
db = connect('MP_100_bgfe_train.db')

# Iterate through entries
for row in db.select():
    atoms = row.toatoms()  # ASE Atoms object
    formation_energy = row.formation_energy
    band_gap = row.band_gap

    print(f'Formula: {atoms.get_chemical_formula()}')
    print(f'Formation Energy: {formation_energy:.3f} eV')
    print(f'Band Gap: {band_gap:.3f} eV')
    break  # remove this if you want to iterate over all entries