Make ES usable with ML-based potentials

[RudolfWeeber]

Goal

• to be able to run a full simulatoin using a machine-learned potential
• The descriptor calculation and ML model evaluation shoudl happen direclty from the C++ core, to obtain good scaleability

Steps

• [X] infrastructure to obtain the neighborhood of a particle
• [ ] Pick an ML-based potential
• [ ] CMake support for the ML potential libraries
• [ ] Identify the information needed for descriptor calculation
• [ ] Implement a kernel for the particle neighborhood providing that information, calling the descriptor and adding the resulting force to the particles
• [ ] Tests/Examples

Technical notes

• If the ML model is a TensorFlow model, it would be good to look into cppflow or other wrapper libraries to avoid making a full tensorFlow installaoitn a dependency of ES

• Relevnat functions in src/core/cells.cpp

  • get_interacting_neighbors() for a given particle returns a vector of pointers to neighboring particles (based on distance cutoff). The resulting vector of Particle* can be used to calculate the descriptor for the particle around which the neighborhood was calculated
  • get_neighbor_pids() For each particle on that MPI rank returns the particle ids of the surrouding partilces. This can be taken as an example/starting point to calculate descriptors for many particles.
• Further optional improvement is possible. E.g., it might be useful to run the actual ML model in a microsevice and call it via RPC from ES. This would allow to chose resource sfor the ML model independently from the ES paralleliziton. This is particularly important, because the local domain size handled by an individual MPI rank in ES cannot be smaller than the search radius in the neighborhood calculation.

[RudolfWeeber]

This ticket is suitable for a new developer, but it may be helpful to read up on lambda with capture expressions in C++, before starting.

[SamTov]

As discussed internally, we will move forward with the ML package from the Csanyi group, MACE:

https://github.com/ACEsuit/mace

This comment can be elaborated upon after our meetings with the group in regard to what can be implemented on the engine side.