GRACE-OFF: A machine-learned interatomic potential for organic liquids using the GRACE architecture

We provide the follwing files: input.yaml: example input script for fitting the 2-layer medium GRACE-OFF model SPICE_V2.0_cleaned.xyz: contains all structures with their corresponding energies and forces, extracted from the original SPICE v2.0 repository (https://github.com/openmm/spice-dataset/tree/main/downloader). This dataset forms the basis for the training, validation, and testing splits used in this work. training_set.pkl.gz and test_set.pkl.gz: contain the training and validation splits used for training the GRACE-OFF models with gracemaker. GRACE_test_set.xyz: contains the 10,000 structures that were excluded from the dataset prior to training and validation of the GRACE-OFF models. These structures were not seen during GRACE training or validation. However, it cannot be excluded that some of them were used during training of the MACE reference models. biased_test_set.xyz: contains 10,000 structures randomly sampled from SPICE_V2.0_cleaned.xyz. These structures may have been included in the training or validation data of both the GRACE-OFF and MACE-OFF models.