Feature: PLUMED interface for the MD driver

[MarkYoungIV]

Summary

This PR adds PLUMED support to ABACUS molecular dynamics: collective variables, biasing potentials (metadynamics, walls, umbrella sampling, ...) and the PLUMED post-processing workflow can now be used together with calculation md. Feedback on design and scope is very welcome.

Usage

plumed                 1
plumed_file            plumed.dat

with a standard PLUMED input file, e.g. well-tempered metadynamics along a distance:

UNITS LENGTH=A TIME=fs ENERGY=eV
d: DISTANCE ATOMS=1,2
metad: METAD ARG=d PACE=100 HEIGHT=0.05 SIGMA=0.1 BIASFACTOR=10 TEMP=300 FILE=HILLS
PRINT ARG=d,metad.bias FILE=COLVAR STRIDE=10

What is included

• new module source/source_plumed/ wrapping the PLUMED C API (plumed_create / plumed_cmd / plumed_finalize), driven from the shared MD loop in source_md/run_md.cpp, so all integrators (md_type) are covered;
• two new INPUT items: plumed (Boolean) and plumed_file (String);
• optional build support: cmake -DPLUMED_ROOT=<plumed-install>; when unset, nothing changes and PLUMED is not required at all;
• documentation: docs/parameters.yaml and docs/advanced/input_files/input-main.md regenerated with the branch binary (--generate-parameters-yaml), plus a new page docs/advanced/interface/plumed.md.

Design notes

• The ABACUS atomic units (Bohr, Hartree, atomic time unit) are declared to PLUMED through setMD*Units; positions, cell, forces, potential energy and masses are exchanged in native units and PLUMED converts them internally, so no manual conversion is needed.
• The call sequence follows the reference implementations shipped with PLUMED (setPositions/setBox/setEnergy -> prepareCalc -> setForces/setVirial -> performCalc); the virial is handed over whenever the MD provides it.
• Every MD step is flagged as a checkpointing step, so PLUMED flushes COLVAR/HILLS continuously and the run can be followed live.
• Current limitation: a single MPI rank is required for the MD part (the atoms owned by the master rank are handed over to PLUMED). This covers typical serial/OpenMP production runs; multi-rank support would require gathering the distributed atoms, which is left for a future PR.

Testing

• Builds: (a) a workstation build with GCC + CUDA (GPU, cusolver) and (b) an HPC CPU cluster build (GCC toolchain, OpenMP + ELPA, ks_solver genelpa), both against PLUMED 2.10.1.
• Validation: LCAO MD metadynamics of the ethane torsion (following the ABACUS example of the DP-GEN tutorial): the COLVAR values are bitwise identical between the plain CPU, the cluster (OpenMP/ELPA) and the GPU builds; the reconstructed FES reproduces the three-fold torsion profile with a barrier of ~12 kJ/mol.
• Application: metadynamics of HNO3 dissociation on a graphene flake at 600 K (3 coordination-number CVs + a spherical wall; 70 atoms, gamma-only LCAO) - the interface has been running for >10k MD steps with stable SCF and continuous HILLS/COLVAR output.
• Documentation CI: parameters.yaml and input-main.md were regenerated with the branch binary and are consistent with the source.

Notes for reviewers

• The feature is compile-gated (-DPLUMED_ROOT), so all existing CI configurations are unaffected.
• Automated tests were not added because the CI builds do not provide a PLUMED installation; the feature was validated manually as described above.
• Happy to adjust naming/placement (e.g. INPUT keyword names) to project conventions.

[MarkYoungIV]

Thanks for the guidance — closing this PR as advised. Two compliance issues:

1. the submission should target deepmodeling/abacus-develop (the active development repo), not this one;
2. the new top-level source/source_plumed/ folder violates the no-new-folder policy under source/.

I will rebase onto the latest deepmodeling/abacus-develop:develop, move the PLUMED interface into source/source_md/ (it is driven from the MD loop), rebuild and re-verify on our CPU cluster, and open a fresh PR against the correct repository. Sorry for the noise, and thanks for the review!