V 1.1 features

[ipcamit]

A public list of upcoming features for version 1.1. Keeping it here for visibility, recommendations and tracking!

1. MIC graphs [high priority/ low effort]
2. Dataset exporting [high priority/ low effort]
3. Dataset.from_lmdb for loading lmdb dynamic datasets [high priority/ low effort]
4. Per atom weights for GNNs [high priority/ high effort]
5. EAM and SNAP Trainer [high priority/ high effort]
6. Documentation update: i. TorchML v001, ii. GNN from KLAY (linked to KLAY v1.0) [high priority/ low effort] (KLAY-KLIFF documentation for a MACE model here: https://klay.readthedocs.io/en/latest/kliff.html )
7. MCMC trainer subsuming the UQ module, interfacing with PyMC3 or Pyro [low priority/ high effort]
8. KLIFF analyzer based off Josh's LTAU

[mjwen]

Given that we have the general KIM trainer, I believe training EAM is currently supported, right? Of course, we might not want to optimize all the knots in the tabulation.

[mjwen]

For 2. dataset exporting to files, what format you are thinking of?

[ipcamit]

I believe training EAM is currently supported, right?

Technically yes. But practically when training the EAM potential using current trainer, scipy try to optimize all of the tabulated values. So that results in optimization over 10000 params or so and scipy gets stuck. So I have a version right now that Pamela and Ellad are evaluating which samples user defined number of knots to optimize, and save it as valid KIM model afterwards (EAM save_kim_model function does not work either). Currently it only supports FS type potential.

For 2. dataset exporting to files, what format you are thinking of?

Something complementary to from_* functions. i.e. each input type will have corresponding dataset output type. I believe it will be i) KLIFF specific xyz, ii) ASE kind of xyz, iii) write to colabfit, iv) lmdb.