{}
Update
As of the mlip v0.1.2 release, the SPICE2 curated dataset was also updated. The main reason the removal of a small number of structures that led to unusually high energy errors with our models. We also wanted to remove structures with Li, K and Na atomis, as these elements were absent from the validation set. 0.2 % of the training set and validation set was removed. Details added to the filtering process below.
SPICE2_curated
The provided dataset is based on the version 2 SPICE dataset. The SPICE dataset was chosen for its diversity, both in chemical and conformational space, compromising of approximately 2 million structures computed at the ωB97M-D3(BJ)/def2-TZVPPD level of theory.
Dataset Structure
The provided dataset is split into a training set and a validation set with a 95/5 split. The split was done per molecular SMILES, insuring that different conformations of the same molecule do not appear in both the training and validation set. The training set contains 1 734 158 structures and the validation set 87 764 structures.
Filtering Process
The version 2 SPICE dataset was filtered by removing the following structures:
- Unphysical structures: Only keeping structures where all hydrogen atoms have exactly one bond (removing 42 689 structures).
- Charged systems: Removing all charged systems (removing 142 647 structures).
- High forces: Applying a total force filter of 0.1 eV/A and a maximum force filter of 15 eV/A (removing 1 024 structures).
- High energy errors: Removing structures with high z-score when computing energy errors with our three trained models (removing 3656 structures from training and 158 from validation).
- Atomic species Removing any structures containing Li, K or Na (removing 82 structures)