Evaluate vdW corrections once per ionic step by Growl1234 · Pull Request #7712 · deepmodeling/abacus-develop

Growl1234 · 2026-07-30T13:19:23Z

Previously, ABACUS evaluated the vdW energy before SCF and then created additional vdW solver instances in the force and stress code paths. Depending on the calculation type, the same dispersion correction could therefore be calculated two or three times within one ionic step. This affected PW, LCAO, OFDFT, and DoubleXC workflows. The duplicated work was particularly noticeable for DFT-D4, where calculations could pause both before and after SCF.

This PR refactor the vdW correction workflow so that the required dispersion quantities are evaluated only once for each ionic configuration.

Closes #7680.

Changes

Introduce explicit VdwRequest and VdwResult types.
Evaluate the vdW correction once in ESolver_FP::before_scf().
Store the resulting energy, forces, and stress for the current ionic step.
Pass the cached result explicitly to PW, LCAO, and OFDFT force/stress assembly code.
Remove vdW solver construction from downstream force and stress modules.
Reuse the same vdW result in the DoubleXC workflow.
Remove the previous getter interfaces that could implicitly trigger additional calculations.

The individual vdW implementations were also adjusted so that all requested quantities are produced by one backend evaluation:

DFT-D2 accumulates energy, forces, and stress in one pair/image traversal.
DFT-D3 accumulates energy during the existing force/stress pair and ATM traversals, avoiding a separate energy evaluation.
DFT-D4 obtains energy, gradient, and virial from one dftd4_get_dispersion() call.

For energy-only calculations, the backends retain an energy-only path and avoid unnecessary derivative calculations.

DFT-D3 considerations

The DFT-D3 refactoring requires particular care because its energy and derivative implementations previously followed partially separate paths.

The combined implementation preserves the existing periodic-image counting factors for pair and three-body ATM contributions, including self-image and repeated-index cases. Coordination-number derivatives and the existing force and stress assembly remain unchanged.

Testing

Here is my simple test: test.zip

In the tested cell-relaxation case, the relevant timings changed from:

Vdwd4::cal_energy: 4.82 s [Pre-SCF]
Vdwd4::cal_force: 17.03 s [Post-SCF]

to:

Vdwd4::evaluate: 15.89 s [Pre-SCF]

The total DFT-D4 time was therefore reduced from approximately 21.85 s to 15.89 s, while the total runtime decreased from 49.79 s to 43.16 s.

The pre-SCF DFT-D4 timing apparently increased, which exactly reflects that all requested dispersion quantities are now calculated and accounted for at that point, rather than being split between the pre-SCF and force-assembly stages.

Copilot