Repository navigation
Missing boundary_shift updates in LCAO RT-TDDFT MD #8221
Description
Thanks to @Critsium-xy for spotting this during the review of #8207.
The new MD path wraps atomic coordinates and synchronizes them to UnitCell, but does not update boundary_shift. This can cause reset_matrix_boundary() to skip the phase corrections needed when atoms cross periodic boundaries at non-Gamma k points.
This is separate from the sub-roundoff wrapping fix in #8207. We should restore the old boundary_shift behavior and add a minimal regression test with a finite displacement across a periodic boundary.
Activity
ABACUS-PR-BOT commentedon Oct 27, 2026
Confirmed by source inspection at
00d74be8e24cd9e7ab3537c8088f8bada783ab4c:MD_base::update_pos()discards the integer translation during wrapping, and neitherMDCell::sync_backing_unitcell()branch updatesboundary_shift. The LCAO RT-TDDFT runner still uses that field to rephase saved H, S and coefficients, so a genuine finite crossing leaves the correction inactive. The MD transition in #7812 removed the old producer call; #8207's rounding clamp does not restore it.A regression should exercise the complete drift → backing-cell sync → phase-correction path, using two atoms with nonzero inter-atom matrix elements, finite crossings in both directions, and k=(1/4,0,0). It should also check a subsequent noncrossing step and non-root/empty MPI ownership. Coordinate-only or Gamma-only checks can miss this omission.
This is a source-supported diagnosis, not an executed reproduction of an energy jump. No product changes were made. Restart/substep behavior, nonzero hybrid-gauge phases and gathered-matrix layouts still need focused verification before claiming a complete repair.