Current Posture on Model Interface Additions and Support

[orionarcher]

We want TorchSim to grow quickly and support a wide range of models from across the community. We also recognize that the MLIP ecosystem is evolving quickly: packages push breaking changes, dependencies evolve and conflict, and interfaces change.

We imagine three possibilities for model support:

1. Internal Models: All Model Interfaces are defined within TorchSim and maintained by the TorchSim team or the model creators. The approach taken by a package like graph-pes.
2. External Models: All Model Interfaces are defined outside TorchSim and maintained by the model maintainers. The ASE Calculator model of support.
3. Mixed Support: A mix of both.

We intend to start with the Internal Models approach. We believe it is the best way to maintain a consistent interface, detect errors quickly, and create an easy developer experience. It works especially well if model developers are able to maintain and update their own interfaces. However, if we find that interfaces are regularly breaking, environmental configurations are unstable, or subtle errors are being introduced, we may offload some Model Interfaces to the MLIP packages themselves (approach 2/3).

We will update this discussion as our thinking evolves.

[CompRhys]

As the Nequip framework broke the dam and chose to in-house their TorchSim interface for greater control and optimization, we have decided that the package has been deemed "mature" enough by the community to move towards the External Model Posture. Whilst we are not going to accept new implementations in this repo current policy is that we will continue to accept PRs that wrap external interfaces under the External posture.