SPICE2_curated_v2

Following the mlip v0.2.0 release, we provide here a new SPICE curated dataset, used to train new pre-trained models. This dataset is based on the SPICE subset of the OMOL25 dataset, comprising approximately 2 million structures computed at the ωB97M-V/def2-TZVPD level of theory. When compared to the data provided with the mlip v0.1.2 release linked here, this version introduces the following modifications:

• Updated data source: Filtering is now conducted on the SPICE subset of OMOL25, replacing the original SPICE dataset.
• Charged systems: The dataset now explicitly includes charged systems.
• Expanded chemical coverage: The number of covered elements has grown from 12 to 17, through the addition of Ca, K, Li, Mg and Na.

Dataset Structure

The provided dataset is split into a training set, validation set and test set with a 90/9/1 split. The split was done per molecular SMILES, ensuring that different conformations of the same molecule do not appear across sets. The final training set contains 1,763,962 structures covering 17 chemical elements (B, Br, C, Ca, Cl, F, H, I, K, Li, Mg, N, Na, O, P, S, Si), the validation set 172,838 structures across 17 elements, while the test set contains 17,477 structures across 13 elements (natural result of the per-SMILES split).

Filtering Process

The SPICE subset of OMOL25 was filtered by removing the following structures:

• Unphysical structures: Only keeping structures where all hydrogen atoms have exactly one bond.
• High forces: Applying a total force filter of 0.1 eV/Å and a maximum force filter of 15 eV/Å.

OMOL25 dataset source

https://doi.org/10.48550/arXiv.2505.08762