Planning for the transition to metatomic

[Luthaf]

The code for atomistic models used to live in https://github.com/metatensor/metatensor, as part of the metatensor-torch package. We would like to separate it into a separate package (working name metatomic) to make the separation between the data storage and atomistic interface clearer.

This will also be the opportunity to make the model interface a bit more abstract and not reliant on TorchScript. TorchScript models will still be supported, but will only be one option among others (including e.g. pure python models or pure C++ models).

To minimize churn and user pain, here is a plan for the transition:

Phase 1: bring the new git repository up to speed

• integrate with codecov for code coverage
• make a first release of the metatomic and metatomic-torch Python packages
• move issues from the metatensor repository

I'm not sure about the following: doing it early might help smooth the transition down the line, but could have unforseen issues with TorchScript

• rename classes (MetatensorAtomisticModel => MetatomicModel; NeighborListOptions => PairListOptions; …). We will want to do this at some point anyway, and if we can re-export them with a different name in metatensor.torch.atomistic that should make the transition smoother (models exported with metatensor.torch.atomistic will work with metatomic).

Phase 3: Update engine to use metatomic

Before we can have model rely on the metatomic code, we will need to update all the simulation engine to use the new metatomic library

• LAMMPS
• PLUMED
• i-PI
• chemiscope

This update would still use the TorchScript version of the models directly, abstracting away from this will be done later down the line.

Phase 2: re-export metatomic.torch from metatensor.torch.atomistic

• remove the code from metatensor-torch, and add a dependency on metatomic-torch there. Then metatensor.torch.atomistic can re-export the classes from metatomic.torch; giving source compatibility to most of the existing code. This re-export should also have a deprecation warning pointing people to use metatomic instead

Phase 4: Remove from metatensor-torch

• After a while, remove the deprecated code from metatensor-torch

Not part of the transition, but required anyway

• write developer documentation