[Growl1234]

Follow up to #8069. Dropped the manual configuration from CI.

[Growl1234]

I'll take a look at the testing failure tomorrow. nan is weird.

[ABACUS-PR-BOT]

Static review of 615ad580a2a0 against 8e2017587f62. No builds or runtime tests were run; CI diagnostics were unavailable. The alternative-version suggestion is nonblocking and concerns an inherited limitation.

[ABACUS-PR-BOT]

[P2] Retain tracking-checker regression coverage

Please keep the test_es_check.py invocation when removing the custom LibXC setup. This script is not registered with CTest, so the remaining steps no longer run its negative cases for NaN forces, lost branches, and missing tracking diagnostics. Restoring this step, or registering the script with CTest, preserves that coverage.

[ABACUS-PR-BOT]

[P3] Cover the alternative LibXC version in a follow-up

Non-blocking: --package-version libxc:alt selects 7.0.0, whose build ignores MAXORDER and defaults to DISABLE_KXC=ON, so that installation still lacks KXC for LR gradients. This is an inherited limitation; the default 7.1.2 path is covered by this change. A follow-up after merge can use version-appropriate derivative switches or document that KXC currently requires the main version.

[maki49]

My local test finds LibXC 7.0.0 pass three isolated LibxcSanitizer tests, including GgaSigmaReverseMatchesTheWeightedEnergy, which fails with 7.1.2 producing NANs. Both the CI input (0.04, -0.20, 0.03) and the physically valid opposing-gradient input (0.04, -0.04, 0.04) produce finite energy and first derivatives in 7.0.0.

The suspected regression is in PBE correlation’s generated evaluation. Commit b5a0dec introduces sqrt(q), where q = sigma_uu + 2*sigma_ud + sigma_dd, together with 1/sqrt(q) in the first-derivative chain; its immediate parent uses q directly. I performed a runtime bisect and this has been confirmed:

LibXC commit	Standalone probe: eight inputs	ABACUS test GgaSigmaReverseMatchesTheWeightedEnergy
Parent 321b7a9	All outputs finite	pass
b5a0dec, “Regenerated sources”	Cases 3 and 5 produce NaNs	fails

The eight inputs in the standalone test are listed as follows. Sigma components are ordered as (uu, ud, dd). “Finite” means that exc, both vrho components and all three vsigma components are finite.

Case	sigma_uu	sigma_ud	sigma_dd	Parent 321b7a9	Commit b5a0dec
0	0.04	0.01	0.03	Finite	Finite
1	0.00005	0	0.03	Finite	Finite
2	0.04	0.20	0.03	Finite	Finite
3	0.04	−0.20	0.03	Finite	exc, vrho, vsigma are NaN
4	0.00005	0.20	0.03	Finite	Finite
5	0.04	−0.04	0.04	Finite	exc and vrho finite; vsigma NaN
6	0.04	−0.034	0.03	Finite	Finite
7	0.04	-sqrt(0.04*0.03) ≈ −0.034641016151377546	0.03	Finite	Finite

For case 3, internal cross-sigma clipping leaves q = −6.9388939039072284e−18 through floating-point rounding, making the square root invalid. For case 5, q=0 exactly; the energy remains finite, but the derivative chain encounters an unresolved zero-times-infinity limit.
Importantly, case 5 is physically valid: equal and opposite spin-density gradients give zero total gradient.

The probe calls xc_gga_exc_vxc() for XC_GGA_C_PBE (ID 130), with:

xc_func_init(&func, XC_GGA_C_PBE, XC_POLARIZED);
xc_func_set_dens_threshold(&func, 1e-6);
xc_func_set_sigma_threshold(&func, 1e-2);

double rho[2] = {0.4, 0.2};

The similar issue has been reported in https://gitlab.com/libxc/libxc/-/work_items/639 and fixed in its devel branch. 7.1.0-7.1.2 all includes this regression. We can use 7.0.0 by default or wait for the upcoming 7.2.0 release.

[Growl1234]

Thanks for the reminder! Then we can wait for the 7.2.0 update. Version 7.0.0 takes much more time building with KXC.

[AsTonyshment]

I encountered exactly the same NaN issue with Libxc v7.1.2 a while ago, and discovered that the $\frac{1}{\sqrt{q}}$ issue had already been fixed in MR #839 just as I was about to open an issue in the Libxc repository. They were pretty quick to fix this one, considering that a previous issue I reported took about two years to get resolved😅