metadata
license: cc-by-4.0
These are ALIGNNd models trained with standard quantile loss to predict the kpoints-density for different quantiles. For each quantile top 3 (quantile loss minimal on the validation set) checkpoints are recorded.
The implementation of ALIGNNd model can be found here https://github.com/stfc/goldilocks_kpoints. For these checkpoints input features are embeddings/atom_init_with_sssp_cutoffs.json, additional features are composition, structure, lattice, and metallicity embeddings
Performance of the model trained for 0.5 quantile is:
MAE: 0.069
MAPE: 0.189
MSE: 0.0097
R2 score: 0.697
Spearman_corr: 0.866
Kendall_corr: 0.677