Feature: point-counter-charge correction by LKFEIYI · Pull Request #8075 · deepmodeling/abacus-develop

LKFEIYI · 2026-10-04T23:53:05Z

What's changed?

Add assume_isolated pcc_0d for molecules in cubic cells and assume_isolated pcc_2d for periodic slabs. The correction energy is reported as E_pcc; pcc_2d_axis selects the open lattice vector (0/1/2, default 2).

This is the PCC-only part of #8052. Geometry and atom-data extraction are placed in source_cell, FFT-grid coordinates in source_basis/module_pw, and charge moments and PCC kernels in source_estate alongside Makov–Payne.

PCC0D requires a cubic cell. PCC2D requires the open vector perpendicular to the periodic plane, Gamma-only sampling along it, and symmetry operations that preserve the open direction. The wrapping boundary must lie in vacuum. PCC2D uses the open planar monopole coefficient -pi*L/(3*A).

Linked Issue

Follow-up to #8052, split according to the review feedback.

Build: cmake --build build_pcc0d --target cell MODULE_CELL_cell_geometry MODULE_CELL_cell_tools --parallel 4 passed. CTest: both focused targets passed (10 tests), OMP_NUM_THREADS=1 outside sandbox. C++11 syntax checks passed for both changed production sources. Governance: no blockers; geometry header needs Vector3 and array for value members, string for validation diagnostics, and vector for input containers. No INPUT or user-visible behavior changes in this foundation step.

Place charge moments and PCC formulas beside makov_payne; geometry and atom extraction remain in source_cell. No INPUT behavior changes yet. Built both focused targets and passed 11 tests outside sandbox with OMP_NUM_THREADS=1, including energy/potential and ionic-force finite differences. C++11 syntax checks passed. Governance has no blockers; includes define the value moment type, diagnostic string and kernel interface.

Register standalone PCC for CPU KS-DFT PW/LCAO, add its energy once, and expose force and electrostatic output from the same component result. Generated INPUT YAML/Markdown from the new CLI. Full abacus_basic_para build and focused potential, energy, INPUT and grid tests passed. Neutral water, charged H3O+, LCAO and two-rank MPI SCF runs converged; old H3O+ energy/forces match at printed precision. C++11 syntax and staged governance checks passed without blockers. Remaining PARAM read is the existing PW registration dependency (net zero); header includes provide value member types, base class and standard containers/diagnostics.

Keep potential update orchestration distinct from geometry/atom preparation and distributed density integration. Rebuilt abacus_basic_para, MODULE_ESTATE_pot_pcc and MODULE_ESTATE_potentials_new; both CTest targets passed outside sandbox with OMP_NUM_THREADS=1. Quality score improves from 82 to 90. No INPUT behavior changes.

Register neutral water and charged H3O+ PW cases plus neutral water LCAO. Collect E_pcc in eV alongside existing total-energy and force properties. The new references use two MPI ranks; existing references are unchanged. Verification: - cmake --build build_pcc0d --target abacus_basic_para --parallel 2: passed. - OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 Autotest.sh, -n 2 -o 1: PW 220_PW_PCC0D_WATER / 221_PW_PCC0D_H3O and LCAO scf_pcc0d_water / existing scf_solvation passed default thresholds. - Nine targeted CTest executables passed, including INPUT and PotPcc. - Staged governance and git diff --cached --check: passed.

Drop the newly added 220_PW_PCC0D_WATER case and CASES_CPU entry. This case took about 48 seconds and overlaps neutral LCAO plus analytic and host coverage. Retain charged H3O+ PW and neutral water LCAO so both solver paths remain covered. Existing tests and thresholds are unchanged. Verification (OMP_NUM_THREADS=1, OPENBLAS_NUM_THREADS=1, outside sandbox): - Autotest.sh -n 2 -o 1 -j 1 -r '^221_PW_PCC0D_H3O$': passed, 10.826 s; energy, force and E_pcc checked at default thresholds. - Autotest.sh -n 2 -o 1 -j 1 -r '^scf_pcc0d_water$': passed, 9.858 s; energy, force and E_pcc checked at default thresholds. - Nine targeted CTest executables: passed, 0.21 s total. - Staged governance and git diff --cached --check: passed. Only test files and registration changed; no executable rebuild required.

Build: abacus_basic_para, MODULE_CELL_cell_geometry and MODULE_ESTATE_pcc_2d passed. Both focused CTest targets and explicit C++11 syntax checks passed. This stage adds no INPUT behavior; user documentation follows in the integration stage. pcc_2d.h includes the charge-moment API used by its public kernel declarations; cell geometry is forward declared.

Keep dimension and open axis immutable at construction. Reuse charge moments, pool reductions, energy, force and electrostatic output consumers. Require Gamma-only open-direction sampling and symmetry that preserves the slab. Add pcc_2d_axis INPUT metadata and regenerate both parameter documents. Report the actual failure stage and reject missing density storage via the standard ABACUS WARNING_QUIT path. INPUT errors identify conflicting settings. Verification: - Full executable and affected unit targets built successfully. - Ten targeted CTest targets passed in 0.22 s (OMP_NUM_THREADS=1). - Explicit C++11 syntax checks passed for geometry, kernel and PotPcc. - PW charged and LCAO neutral 2D SCFs converged on two MPI ranks; comparison with the saved original executable agrees within 2.7e-10 eV energy and 2.5e-5 eV/Angstrom force components (old run used one rank). - Metadata generation and pcc_2d_axis CLI help passed. - Staged governance and diff checks passed without blocking findings. Global references: added 2, removed 6, net -4. The existing potential factory reads the axis from PARAM and passes it explicitly into PotPcc; PW setup now shares one const INPUT reference for existing component selection. PotPcc and numerical helpers add no globals. The new header include owns Pcc2dParameters by value and thus needs the complete definition.

Add only two small vacuum PCC 2D integration cases: charged H3O+ PW, open axis 2, and neutral water LCAO, open axis 1. Generate references with two MPI ranks and one OpenMP thread. Keep all old references and thresholds unchanged. Verification with OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1: - PW Autotest.sh -n 2 -o 1 -j 2 selecting both PCC0D/PCC2D H3O+: passed. - LCAO Autotest.sh -n 2 -o 1 -j 2 selecting both PCC0D/PCC2D water: passed. - New tests compare total energy, force magnitude and E_pcc. - Staged governance and git diff --cached --check: passed. - Two-rank runtime rejection of tilted open vector and non-Gamma open sampling passed; kpar=2 invalid sampling also exits 1 without hanging. Only calculation inputs and registration changed; no rebuild needed.

Restore assume_isolated and pcc_2d_axis descriptions from abacus_sccs, removing only solvent clauses. Regenerate parameter YAML and input-main. Remove the PCC solvent-combination test, solvent wording in the added energy test name/readmes, and explicit imp_sol=0 from four PCC cases. Keep the INPUT mutual-exclusion check as requested; physics is unchanged. Verification: - Full executable, INPUT test and energy test rebuilt successfully. - Both affected CTest targets passed in 0.18 s with OMP_NUM_THREADS=1. - CLI parameter help and --check-input passed for all four PCC cases. - Script comparison confirms exact original descriptions except the removed solvent clauses and confirms the mutual-exclusion check remains. - Generated docs contain 567 parameters. - Staged governance and git diff --cached --check passed.

mohanchen · 2026-10-05T04:11:25Z

Thanks! This is better, we will review it soon.

kluonju

Overall I think the tests need some more thoughts to reflect close-to real scenarios.

kluonju · 2026-10-06T03:05:34Z

+namespace elecstate
+{
+
+/// Self-consistent vacuum PCC. Owns the result of its last density update.


It would be good to know the reference equations that is used for this implementation. A paper mentioned below? O. Andreussi and N. Marzari, Phys. Rev. B 90, 245101 (2014) or some ref to other codes.

Added. The implementation mainly follows ENVIRON. The one difference is the monopole constant for charged slabs. Andreussi and Marzari, Phys. Rev. B 90, 245101 (2014), and QE-Environ use −π/(3L), where L is the cell length along the open axis. We use −πL/(3A), with A the periodic area, following the reminder from Critsium-xy in #8052.

It (Fig. c) shows that the coefficient -pi * L_y / (3 * A) works.

kluonju · 2026-10-06T05:56:14Z


 namespace elecstate
 {
+const PotPcc* Potential::pcc_component() const


adjust the header definition and consider to adapt these public functions as private member functions of class Potential. These lines should not be placed ahead of the constructor.

Thanks for the suggestion. I moved the definitions after the constructor/destructor in potential_new.cpp, grouped the declarations with the other public getters, and moved pcc_component() to the private member functions.

kluonju · 2026-10-06T05:58:27Z

 bool XC_Functional::ked_flag = false;
 namespace elecstate
 {
+double Potential::pcc_energy_rydberg() const


This test seems unnecessary.

The remaining Potential::pcc_energy_rydberg() definition in this file is not a test but a link stub. It gives the linker a definition for a Potential member that ElecState::cal_energies references; here, the call is this->pot->pcc_energy_rydberg() in the pcc_0d/pcc_2d branch. It returns 0.0 and is never reached by the remaining tests, which all use assume_isolated = "none". Removing it would require either linking potential_new.cpp and its dependencies into this lightweight test or changing the cal_energies interface.

kluonju · 2026-10-06T06:00:42Z

    EXPECT_DOUBLE_EQ(elecstate->f_en.etot, 1.3);
 }

+TEST_F(ElecStateEnergyTest, AddsPccEnergyOnce)


Also this test is not needed. Consider to add more meaningful ones of an actual model or something else.

removed. Four new tests were added.

- Document the PCC 0D/2D kernels, their references (Andreussi and Marzari 2014, Dabo et al. 2008, Rutter 2021 for the charged-slab constant) and the ENVIRON implementation they follow in pot_pcc.h; cite Rutter 2021 next to the 2D gauge in pcc_2d.h. - Define the Potential PCC accessors after the constructor/destructor, declare them with the other public getters, and keep the pcc_component() helper with the private member functions. - Drop the stubbed-energy PCC test from elecstate_energy_test.cpp. - Add physical checks: periodic Hartree energy + E_pcc recovers the isolated energy of a charged Gaussian ion and of a polar Gaussian pair (0D), and of a charged layer and a polar neutral slab (2D), independently of the cell size. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141raGGLbA6qqg3srvRcPnz

The geometry, grid and moment helpers only receive values that ABACUS has already built from STRU and INPUT, so their finite/size/storage checks and bool+error returns could never fire. - grid_positions and charge_moments return their result directly. - weighted_center returns the center; make_pcc_2d_parameters returns the parameters. - make_orthogonal_cell, make_slab_cell and make_pcc_0d_parameters keep a bool for the structure conditions a user can violate (perpendicular, cubic); PotPcc turns them into WARNING_QUIT messages. - PotPcc no longer reduces replicated structure checks over the pool (every rank sees the same cell and stops together), and drops the storage, nonfinite-result and stale-result checks. validate_kpoints keeps its cross-pool reduction, since each pool holds its own k points. - write_elecstat_pot drops the PCC potential size check. Tests for the removed failure paths are removed; a death test covers the non-cubic 0D and tilted 2D cells. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BKqk2SNWXoZn2PRRGRzGca

PotPcc takes the expected electron count (INPUT nelec) and writes a warning when the PCC net charge differs from it by more than 1e-6 e, e.g. after a very tight diagonalization. The correction keeps using the grid charge, as before.

Drop the symmetry and origin details from assume_isolated pcc_0d and pcc_2d and the LCAO load-balance note from pcc_2d_axis; the run still stops with a clear message when the cell or symmetry is unsupported.

The exact monopole and dipole values are implied by the isolated-energy tests, and the bilinear kernel is an internal step of pcc_0d_energy. The finite-difference checks of each kernel, the orthogonal and slab geometry checks, the PotPcc input rejections and the charge-moment checks now share one test each. Every property stays checked; 31 tests become 18.

LKFEIYI added 12 commits October 4, 2026 09:47

test: reduce PCC integration cutoffs and SCF cost

d820103

test(pcc): tighten SCF convergence for stable MPI regressions

cd61d43

mohanchen closed this Oct 5, 2026

mohanchen reopened this Oct 5, 2026

refactor(pcc): name computed call arguments

2487df4

kluonju reviewed Oct 6, 2026

View reviewed changes

LKFEIYI force-pushed the codex/pcc branch from 195346b to 0fb7990 Compare October 6, 2026 08:24

LKFEIYI and others added 5 commits October 6, 2026 17:53

Merge branch 'develop' into codex/pcc

2442d9e

docs(pcc): shorten the PCC INPUT descriptions

9667b09

Drop the symmetry and origin details from assume_isolated pcc_0d and pcc_2d and the LCAO load-balance note from pcc_2d_axis; the run still stops with a clear message when the cell or symmetry is unsupported.

LKFEIYI mentioned this pull request Oct 9, 2026

Feature: self-consistent continuum solvation (SCCS) Step1: SCCS base #8098

Draft

7 tasks done

Navigation Menu

Repository navigation

Feature: point-counter-charge correction - #8075

Feature: point-counter-charge correction#8075
LKFEIYI wants to merge 19 commits into
deepmodeling:developfrom
LKFEIYI:codex/pcc

LKFEIYI commented Oct 4, 2026

Uh oh!

mohanchen commented Oct 5, 2026

Uh oh!

kluonju left a comment •

edited

Loading

Uh oh!

kluonju Oct 6, 2026

Uh oh!

LKFEIYI Oct 6, 2026

Uh oh!

kluonju Oct 6, 2026

Uh oh!

LKFEIYI Oct 6, 2026

Uh oh!

Uh oh!

kluonju Oct 6, 2026

Uh oh!

LKFEIYI Oct 6, 2026

Uh oh!

kluonju Oct 6, 2026

Uh oh!

LKFEIYI Oct 6, 2026 •

edited

Loading

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

LKFEIYI commented Oct 4, 2026

What's changed?

Linked Issue

Uh oh!

mohanchen commented Oct 5, 2026

Uh oh!

kluonju left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kluonju Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

LKFEIYI Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

kluonju Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

LKFEIYI Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kluonju Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

LKFEIYI Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

kluonju Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

LKFEIYI Oct 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

kluonju left a comment •

edited

Loading

LKFEIYI Oct 6, 2026 •

edited

Loading