Open chrisiacovella opened 1 month ago
This was mostly addressed in PR #245 . This PR removed configurations with high forces (about 1 hatree/bohr, just like done in spice). This also generates a test/full dataset that only contain the final energy minimized configuration, to make something that is very similar to qm9 (but with forces). This can serve as a baseline for also seeing importance of a few steps of optimization, MD generated configurations, etc.
Currently the PhAlkEthOh dataset comes from the OpenFF optimization dataset and contains the entire optimization trajectory for each unique molecule.
It would be good to have a few additional variations of this dataset for exploring various aspects of the different NNPs and how data generation strategies impact efficacy. A few additional "versions" to add for the existing dataset:
Related, I will work on getting additional calculations going using the trajectories generated with GAFF.