Implement a Metatensor adaptor

[jngrad]

Data flow

See Metatensor - Data flow

• ESPResSo needs to export a neighbor list in the form of a vector of distance vectors
• Metatensor sends back the central particle force, pressure tensor, energy, or observables like induced dipole

Metatensor expects distances, forces and energies in standard SI units. One has to set up the ESPResSo system such that e.g. the reduced unit is in Ångstroms, e.g. box_l=[100, 100, 100] means a box size of 100 Å. This can be achieved by designing a convenience function that leverages pint to create an empty ESPResSo system with the correct size, time step and energy units.

For domain decomposition: typically the ML model neighborhood radius is less than the interaction range. Thus we should be able to use metatensor in parallel without the need to tweak the ESPResSo domain decomposition and cell cutoff. One caveat: without LJ interaction the non-bonded cutoff will be zero, so we need to introduce a mechanism in the Metatensor plugin of the system call to adapt the min_cutoff cell system property.

Contact persons from the ML side: @SamTov, @Luthaf, @PicoCentauri

Roadmap

• [ ] add pytorch dependency to the CI environment
• [ ] include Metatensor via CMake FetchContent
  • document in the user guide which CMake options are needed to help find pytorch (see LAMMPS Metatensor docs)
  • Metatensor build instructions
• [ ] write the C++ adaptor
  • adapt std::vector<int> nbhood(System::System const &system, Utils::Vector3d const &pos, double dist) to return a vector of particle distances (can be achieved with std::ranges::transform and a Particle to Vector3d projector)
  • use pytorch's autograd feature to compute the force as the gradient of the energy
• [ ] introduce a Python API with the following input parameters:
  • ML model and any additional parameters, e.g. file paths
  • list of particle types for which the C++ core needs to export descriptors
    • which data structure would be suitable to represent the primary types and optional cross-terms?
  • figure out how to set system.min_global_cut
  • see Metatensor - Exporting models
• [ ] write unit tests in the testsuite/python folder
• [ ] write a convenience function to instantiate an ESPResSo system with the correct reduced units

See https://github.com/lammps/lammps/pull/4163 for how Metatensor is being integrated into LAMMPS.

[Luthaf]

Metatensor expects distances, forces and energies in standard SI units.

Not quite, but it does have to be one of the named units in there if you want metatensor to perform unit conversion: https://docs.metatensor.org/latest/atomistic/reference/models/index.html#known-quantities-and-units.

If some unit are left unset, metatensor will just pass data as-is from engine to model and respectively (which might fail if the model is not trained to work with the same units as espresso).

The rest of the plan here looks good to me!

[jhossbach]

I will start working on it.