Stabilize RI RPA/EXX and Ewald EXX method merge on develop by AroundPeking · Pull Request #7019 · deepmodeling/abacus-develop

AroundPeking · 2026-03-12T03:30:10Z

Summary

This PR consolidates the RI RPA/EXX merge work onto current develop and resolves the main regressions found during validation.

Main changes

integrate the RI/RPA post-SCF workflow with current develop
restore RI Coulomb-related input handling and output behavior
restore sparse matrix output precision needed by downstream tools
fix nspin=2 RI / overlap-matrix output inconsistencies
fix TRS symmetry-index handling in symrot_k output to avoid symmetry-related crashes
keep the merged behavior consistent with pre-merge ABACUS results for the checked LibRPA workflows

Validation

Checked against the regression cases discussed during debugging, including:

SOC + RPA/SCF workflow
Coulomb cut / mat output consistency
nspin=2 LibRPA-related workflow consistency
symmetry-enabled EXX/RI run that previously crashed in print_symrot_info_k()

Notes

This branch is intentionally squashed to make review easier.
The previous detailed development history is preserved locally in a backup branch if needed.

- integrate the RI/RPA post-SCF workflow with develop - restore Coulomb/input-output handling and sparse matrix precision - fix nspin=2 and TRS symmetry output regressions

The default was accidentally changed to "none", causing all tests without explicit latname to fail at STRU parsing with "do not use LATTICE_VECTORS along with explicit specification of lattice type".

The previous merge accidentally changed exx_singularity_correction reset logic (lc_pbe etc from "spencer" to "massidda") and removed metadata, causing 08_EXX integration tests to fail. Restore the develop version and append new RI parameters (shrink_abfs_pca_thr, shrink_lu_inv_thr, exx_coul_moment, exx_rotate_abfs, exx_multip_moments_threshold, out_unshrinked_v).

The removal of filter_empty_orbs calls caused empty orbital shells to be retained in lcaos and abfs vectors, altering RI exchange integrals for all EXX calculations including standard HF used by LR-TDDFT.

The PR removed the filter_empty_orbs function from Construct_Orbs and inlined its logic into change_orbs() for lcaos. The previous commit incorrectly called the now-deleted function. This commit inlines the same filtering logic directly for abfs to restore the empty orbital shell removal without depending on the deleted function.

…formula - Restore LRI_CV, Matrix_Orbs11/21/22, exx_abfs-construct_orbs, exx_opt_orb, ABFs_Construct-PCA to develop versions to minimize diff - Fix spencer Rcut: use get_nkstot_full() without /nspin0 division (nkstot_full already excludes spin doubling) - Fix shared_ptr<ORB_gaunt_table> usage in Exx_LRI, RPA_LRI - Fix ewald_Vq compilation errors - Fix exx_opt_orb-print.cpp to match develop API

The spencer Rcut formula was corrected to not divide by nspin0 (get_nkstot_full() already excludes spin doubling), which changes the excitation energy result from 6.050430 to 2.574437.

The previous ref (2.574437) was from fisherd which lacks LibXC, producing different results. Use CI environment's actual value.

source/source_cell/k_vector_utils.cpp

mohanchen · 2026-03-13T00:22:56Z

Thanks for your PR, which extends the functionality of ABACUS. I noticed you have introduced several new input parameters. I would suggest adding corresponding unit or integration tests to ensure your results are not affected by future refactoring work from others in ABACUS.

AroundPeking · 2026-03-14T05:26:17Z

Thanks for your PR, which extends the functionality of ABACUS. I noticed you have introduced several new input parameters. I would suggest adding corresponding unit or integration tests to ensure your results are not affected by future refactoring work from others in ABACUS.

Done.

I added integration coverage for the new RPA-related workflow to reduce the risk of future refactoring changing these outputs silently.

The PR now includes:

56_GO_RPA_OUTPUT: a molecular gamma-point RPA case comparing Cs_data_*, coulomb_cut_*, and coulomb_mat_*
57_KP_RPA_SHRINK: a k-point shrink RPA case (1x1x2) comparing Cs_data_*, Cs_shrinked_data_*, coulomb_cut_*, coulomb_mat_*, and shrink_sinvS_*

One follow-up fix was also needed for CI: these RPA files are now compared by absolute value, because the ABF / shrink basis eigenvectors have an arbitrary overall sign, which can flip the sign of Cs_*-type outputs without affecting the physical results when the generated file set is used consistently.

The test additions were added in commit 173490d30, and the absolute-value comparison fix was added in commit 79c17dde6.

mohanchen · 2026-03-15T09:03:49Z

I have some suggestions on the output format of the integration test files in this PR:
The current output format of reference files are not user-friendly. Please add some comment lines starting with # to explain what the data means, so users can understand the content directly by opening the file.
It is not good to print only one number per line. I suggest outputting 6–8 numbers per line for better readability.
The output filenames have redundant words like data, which can be removed. Also, the file index should start from 1 instead of 0.
Some integration test examples produce too many lines (over 10k lines, even >50k lines) and too many files. We should simplify and compress the output properly while keeping the test valid.
These changes will make the test output cleaner, smaller, and easier to check.

mohanchen · 2026-03-15T09:08:28Z

For example, this is the file for DOS:

mohanchen · 2026-03-15T09:09:57Z

This is the file for eigenvalues and occupations:

mohanchen · 2026-03-15T09:11:22Z

This file is for Mulliken charge:

AroundPeking · 2026-03-15T12:46:48Z

I have some suggestions on the output format of the integration test files in this PR: The current output format of reference files are not user-friendly. Please add some comment lines starting with # to explain what the data means, so users can understand the content directly by opening the file. It is not good to print only one number per line. I suggest outputting 6–8 numbers per line for better readability. The output filenames have redundant words like data, which can be removed. Also, the file index should start from 1 instead of 0. Some integration test examples produce too many lines (over 10k lines, even >50k lines) and too many files. We should simplify and compress the output properly while keeping the test valid. These changes will make the test output cleaner, smaller, and easier to check.

Thanks for the suggestions.

For the output format / filename cleanup: I agree this can be improved, but I did not change it in this PR. These RPA output files are part of the current interface used by LibRPA, and the naming/layout is also kept consistent with the existing FHI-aims-style workflow. Changing filenames, indices, or file formatting here would require coordinated downstream changes and is broader than the functional scope of this PR. I would prefer to handle that cleanup in a follow-up PR.

For the integration test size issue, I updated the solid shrink RPA test in this PR. The previous case produced very large reference files, so I replaced it with a much smaller two-atom H 1x1x2 solid case while still checking the same set of RPA-related outputs:

Cs_data_*
Cs_shrinked_data_*
coulomb_cut_*
coulomb_mat_*
shrink_sinvS_*

This keeps the file coverage unchanged but reduces the reference size substantially (from about 224k lines total to 128 lines total).

source/source_lcao/module_ri/LRI_CV_Tools.h

mohanchen · 2026-03-16T04:25:08Z

source/source_lcao/module_ri/LRI_CV_Tools.hpp

this file is strange, you delete 562 lines and add exactly 562 lines, could you double check why?

You are right — these were no-op line-ending / formatting changes from the merge, not real code changes.

I checked the whole PR for the same issue.

Besides the two LRI_CV_Tools files, there was one more no-op formatting / line-ending change in source/source_lcao/module_ri/Inverse_Matrix.h. I removed that as well in commit c33c8d7e1.

I then re-checked the full PR diff against develop, and there are no remaining whitespace-only / line-ending-only file changes.

mohanchen · 2026-03-16T04:28:09Z

I have some suggestions on the output format of the integration test files in this PR: The current output format of reference files are not user-friendly. Please add some comment lines starting with # to explain what the data means, so users can understand the content directly by opening the file. It is not good to print only one number per line. I suggest outputting 6–8 numbers per line for better readability. The output filenames have redundant words like data, which can be removed. Also, the file index should start from 1 instead of 0. Some integration test examples produce too many lines (over 10k lines, even >50k lines) and too many files. We should simplify and compress the output properly while keeping the test valid. These changes will make the test output cleaner, smaller, and easier to check.

Thanks for the suggestions.

For the output format / filename cleanup: I agree this can be improved, but I did not change it in this PR. These RPA output files are part of the current interface used by LibRPA, and the naming/layout is also kept consistent with the existing FHI-aims-style workflow. Changing filenames, indices, or file formatting here would require coordinated downstream changes and is broader than the functional scope of this PR. I would prefer to handle that cleanup in a follow-up PR.

For the integration test size issue, I updated the solid shrink RPA test in this PR. The previous case produced very large reference files, so I replaced it with a much smaller two-atom H 1x1x2 solid case while still checking the same set of RPA-related outputs:

Cs_data_*

Cs_shrinked_data_*

coulomb_cut_*

coulomb_mat_*

shrink_sinvS_*

This keeps the file coverage unchanged but reduces the reference size substantially (from about 224k lines total to 128 lines total).

OK, that sounds reasonable to me.

AroundPeking added 12 commits March 12, 2026 11:24

Stabilize RI RPA/EXX merge on develop

07ccb27

- integrate the RI/RPA post-SCF workflow with develop - restore Coulomb/input-output handling and sparse matrix precision - fix nspin=2 and TRS symmetry output regressions

Fix read_input test build regressions

93ef3df

Fix restart include in input conversion

9c45e4d

Register new input parameters

fe5f2dd

Fix complie errors

6ff343e

fix: restore latname default to "user_defined_lattice"

55ad2b1

The default was accidentally changed to "none", causing all tests without explicit latname to fail at STRU parsing with "do not use LATTICE_VECTORS along with explicit specification of lattice type".

fix: restore filter_empty_orbs in Exx_LRI::init to fix 53_GO_LR_HF test

3215d01

The removal of filter_empty_orbs calls caused empty orbital shells to be retained in lcaos and abfs vectors, altering RI exchange integrals for all EXX calculations including standard HF used by LR-TDDFT.

fix: update 53_GO_LR_HF reference value for corrected spencer Rcut

df20d74

The spencer Rcut formula was corrected to not divide by nspin0 (get_nkstot_full() already excludes spin doubling), which changes the excitation energy result from 6.050430 to 2.574437.

fix: update 53_GO_LR_HF ref to CI-verified value 5.646360

9e6b5a5

The previous ref (2.574437) was from fisherd which lacks LibXC, producing different results. Use CI environment's actual value.

mohanchen reviewed Mar 13, 2026

View reviewed changes

source/source_cell/k_vector_utils.cpp Show resolved Hide resolved

mohanchen added Features Needed The features are indeed needed, and developers should have sophisticated knowledge Refactor Refactor ABACUS codes Input&Output Suitable for coders without knowing too many DFT details labels Mar 13, 2026

AroundPeking force-pushed the pr_merge_develop branch from e643d5c to 9e6b5a5 Compare March 13, 2026 09:51

AroundPeking added 2 commits March 14, 2026 11:18

Fix shrink RPA overlap setup and add regression coverage

173490d

Compare RPA regression files by absolute value

79c17dd

Reduce solid RPA regression output size

bef1621

mohanchen reviewed Mar 16, 2026

View reviewed changes

Drop no-op LRI_CV_Tools formatting changes

09ed2a3

Drop no-op Inverse_Matrix formatting change

c33c8d7

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Stabilize RI RPA/EXX and Ewald EXX method merge on develop#7019

Stabilize RI RPA/EXX and Ewald EXX method merge on develop#7019
AroundPeking wants to merge 17 commits intodeepmodeling:developfrom
AroundPeking:pr_merge_develop

AroundPeking commented Mar 12, 2026

Uh oh!

Uh oh!

mohanchen commented Mar 13, 2026

Uh oh!

AroundPeking commented Mar 14, 2026

Uh oh!

mohanchen commented Mar 15, 2026

Uh oh!

mohanchen commented Mar 15, 2026 •

edited

Loading

Uh oh!

mohanchen commented Mar 15, 2026

Uh oh!

mohanchen commented Mar 15, 2026

Uh oh!

AroundPeking commented Mar 15, 2026

Uh oh!

Uh oh!

mohanchen Mar 16, 2026

Uh oh!

AroundPeking Mar 16, 2026

Uh oh!

mohanchen commented Mar 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AroundPeking commented Mar 12, 2026

Summary

Main changes

Validation

Notes

Uh oh!

Uh oh!

mohanchen commented Mar 13, 2026

Uh oh!

AroundPeking commented Mar 14, 2026

Uh oh!

mohanchen commented Mar 15, 2026

Uh oh!

mohanchen commented Mar 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mohanchen commented Mar 15, 2026

Uh oh!

mohanchen commented Mar 15, 2026

Uh oh!

AroundPeking commented Mar 15, 2026

Uh oh!

Uh oh!

mohanchen Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

AroundPeking Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

mohanchen commented Mar 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

mohanchen commented Mar 15, 2026 •

edited

Loading