Samples with extremely large forces

[Raimondas Galvelis]

The dataset contains samples with forces ~7 magnitudes of order larger than average forces. The force distribution of the SPICE data set:

For reference, the force distribution of the ANI-1x data set:

[Raimondas Galvelis]

I don't see the utility for such extreme samples:

• They will ruin training of NNPs (@peastman is already filtering in his trainings).
• They will ruin benchmarks as majority of NNPs probably won't be fitted for such extremes.

So, people will have to filler. It is guaranteed that it will be done in different ways. So, nobody would be able to say the they used the SPICE dataset without an asterix.

[Peter Eastman]

I'm not sure if it's possible to remove samples from the data on QCArchive? We could add an option to the downloader script to filter them at that point.

These strained samples happened by accident, presumably from differences between the DFT functional and the MD force field used to generate conformations. But in the future we may intentionally add more of them. If we want to be able to train reactive potentials, we'll need to add data where bonds are in the process of being formed. Those will have very high forces. We need to come up with a strategy for handling this.

[John Chodera]

I think the most sensible option is to (1) have the download script add the option to filter (perhaps enabled by default with a sensible threshold), and (2) prepare the default HDF5 file where we've done the sensible thing and removed problematic snapshots with very high forces.

[John Chodera]

#41 changes the behavior of the downloader to filter out samples with large forces by default.