vibe-qcDocumentation

g-xTB v1.1.0: program conventions and departures from the preprint

vibe-qc implements g-xTB natively. The target is the behaviour of the upstream v1.1.0 program; the preprint (Froitzheim, Mueller, Hansen, Grimme, ChemRxiv 2025, doi:10.26434/chemrxiv-2025-bjxvt; “M n” main text and “S n” SI equations) describes the intent. This page records, term by term, where the program does something the preprint does not say or says differently, and the evidence for each statement. It feeds the method paper and the native port.

The reference oracle

  • Program: grimme-lab/g-xtb, tag v1.1.0 (commit df43e0042c6a410fd0873f6605ba88cfb4e2cac9), the Linux x86-64 executable binary/gxtb, identical to the release asset gxtb (59814296 bytes, SHA-256 d6e95e6757c1d2d02ce0e3602188787d4f0c0676ddc15bbd653801dee36a40b4). GPL-3.0; it is run as a separate process and never bundled.

  • Execution: a debian:stable-slim container (linux/amd64 under emulation on an arm64 host), no network, read-only root, one container per job, OMP_NUM_THREADS=1, a fresh work directory per job (the program reads a gxtbrestart file when one is present). Input: a Turbomole coord file in bohr holding exactly the stored coordinates, .CHRG and .UHF; arguments -c coord -molden.

  • Parameter files: the pinned .gxtb, .eeq, .basisq of the same commit (SHA-256 in gxtb_params._SOURCES).

  • Results: tests/data/gxtb_v110_reference.json, written by examples/regression/gxtb/gen_gxtb_reference.py; 59 systems, 176 runs. Two independent generations gave bit-identical total energies. The measurement set tests/data/gxtb_v110_measurement.json (--set measurement) adds free atoms and ions, element probes and scans and open-shell molecules; tests/data/gxtb_v110_spin_ions.json (gen_gxtb_spin_ions.py) free ions of high charge for the spin constants; tests/data/gxtb_v110_stationary.json (gen_gxtb_stationary_states.py) the program restarted from vibe-qc’s converged densities of the degenerate open shells (see SCF).

What the program prints

The term table has 8 decimals: ES1 (charge SIE), ES2+3, ES multipole, ES total, Ex (Mulliken), Espinpol, electronic, atomic core increments, dispersion, nuclear repulsion, total. The energy file holds the total with 14 decimals. molden.input holds the orbitals in the program’s basis (p as x, y, z; d and f as Cartesian functions, each component normalised on its own). Neither H0 nor the ACP energy is printed, and neither is the on-site exchange correction: electronic contains all three.

How terms are isolated

The orbitals of molden.input are orthonormal in vibe-qc’s q-vSZP overlap to 1e-10 for every reference system, so the program’s density matrix P is known in vibe-qc’s basis. Then

E_1 = electronic - ES total - Ex(Mulliken) - Espinpol
    = Tr((H0 + H_ACP) P) + E_OFX

holds to the print precision of the four subtracted terms (2e-8 Eh). The on-site exchange correction E_OFX vanishes for s-only systems and for atoms whose shells are spherically occupied; otherwise it is a linear form in per-element constants that are compiled into the program. Variants of the parameter file switch single terms off (_meta.variants of the reference file): acp_off (ACP coefficients zero), h0_off (H0 levels and their CN slopes zero, so H0 vanishes) and poly_off (element factor of the H0 distance polynomial zero). A test applies the same edit to the parameters it loads from the user’s cache.

Term by term

Basis and charges

  • Preprint: S 26 to S 28 and the EEQ_BC paper.

  • Program: as implemented in gxtb_basis and gxtb_eeqbc: coefficients refer to normalised primitives, each contracted shell is renormalised after the charge-dependent shift, q_eff = q - b q^2 + a sqrt(CN) + c CN q with the .basisq header (a, b, c).

  • Evidence: orthonormality of the program’s orbitals in vibe-qc’s basis to 1e-10 (H to Cl, Ti, I; test test_reference_orbitals_are_orthonormal_in_the_vibeqc_basis).

Coordination number of g-xTB

  • Preprint (S 47): erf counting function with a global steepness and an arithmetic-mean pair radius; the sign of the argument is ambiguous.

  • Program: the stored steepness is negative and is used exactly as printed in S 47, with Rc_AB the mean of the .gxtb element radii. The raw CN enters the H0 levels; the square root enters the repulsion exponent and the ACP weight.

  • Evidence: repulsion parity (below) and the H0 level test with sqrt(CN), which misses by 0.2 Eh.

Atomic increments

  • Preprint (M 13, S 50): a constant per element, zero for H and He.

  • Program: the same.

  • Evidence: printed atomic core increments of all 59 systems to 5e-9 Eh.

Repulsion

  • Preprint (M 14 to 17, S 52 to 56): Z_eff = Z0 (1 - kq q + kq2 q^2) and alpha_AB = alpha_A alpha_B / (alpha_A + alpha_B).

  • Program: Z_eff = Z0 (1 - kq q - kq2 q^2) and alpha_AB = 2 alpha_A alpha_B / (alpha_A + alpha_B); alpha_A = alpha0 (1 + kCN sqrt(CN)); R0_AB the geometric mean; k^pen,1 of H and He for those elements, the heavy value otherwise; exponent 3/2.

  • Evidence: of the eight combinations of the two signs or factors and CN versus sqrt(CN), only this one reproduces the printed repulsion of all 37 multi-atom systems (H, He, Li, C, N, O, F, Na, Cl, Ti, I), to 4.9e-9 Eh; all others miss by 3e-2 Eh or more. A 1 % change of any of the 13 fields moves the repulsion by 1e-4 to 3e-2 Eh.

Extended-Hueckel H0

  • Preprint (M 18 to 21, S 63 to 67): H_l = h_l - k_l CN, off-diagonal (K_l + K_l’)/2 (H_l + H_l’)/2 Pi S~, Pi = pi_lA pi_l’B, pi = 1 + kshp_A kshp_l R/Rc_AB.

  • Program:

    • the stored level is -h_l: H_l = -(lev_l - levcn_l CN), with the raw CN;

    • the Wolfsberg-type factor is not applied to the diagonal, and on-site off-diagonal elements are zero;

    • K_s, K_p and the p factor of the polynomial are stored globals; the s factor of the polynomial is 1; K_d = 1.4, K_f = 2.85 and the d and f polynomial factors (2) are constants of the program, confirmed (see “H0 constants for d and f shells”);

    • the radius of the polynomial is a compiled-in table, measured (below).

  • Evidence: spherically occupied free atoms with the ACP off reproduce sum_l n_l (-lev_l) to 1e-8 Eh (H, He, Li, Be, N, Ne, Na, Mg, Ar, F-, Cl-); all s-only systems of the poly_off variant to 1e-8 Eh; with the polynomial, all s-only systems with one radius per element; the s/p molecules as described below.

H0 basis

  • Preprint (S 1.2, S 1.7.1): exponents scaled shell-wise and k0, k2, k3 of S 28 “replaced” by element values.

  • Program: the three element values multiply the three terms of the SCF-basis effective charge, q’_eff = w0 (q - b q^2) + w2 a sqrt(CN) + w3 c CN q; the exponents are scaled shell-wise; the contraction is renormalised.

  • Evidence: s-only parity to 1e-8 Eh; the literal replacement reading misses H2 by more than 1e-3 Eh, the unmodified SCF effective charge misses the s/p set by 4e-2 Eh.

Diatomic-frame scaled overlap

  • Preprint (S 29 to 38): rotate to the bond frame, scale sigma, pi, delta by 2/(1/k_A + 1/k_B), rotate back; f blocks unscaled.

  • Program: sigma and pi factors from the file as in the preprint (an element without p shells stores 0, which gives a zero harmonic mean, the IEEE limit of the formula); the delta factor is a program constant (5, confirmed: “H0 constants for d and f shells”); f blocks unscaled.

  • Evidence: sigma through every s-only system; pi through CO, H2CO, formamide and the other s/p molecules (below).

Atomic correction potentials

  • Preprint (M 22 to 24, S 76 to 80): theta = c Y_lm exp(-xi r^2) with c = c0 (1 + k CN / CN_avg); H_ACP = S_ACP^T S_ACP, positive semi-definite.

  • Program: H_ACP = sum w <mu|chi><chi|nu> with normalised single-Gaussian projectors r^l exp(-xi r^2) Y_lm for l = s, p, d, f on every atom (16 functions per atom, printed as acpsao), the SCF basis on the AO side, and the signed, linear weight w = c0 (1 + k sqrt(CN)), with no average CN.

  • Evidence: with H0 off, 25 s-only systems and spherical atoms to 1e-8 Eh; the ACP energy of H2 is negative, which no positive semi-definite operator can give; CN in place of sqrt(CN) misses H2 by more than 1e-4 Eh.

The s/p molecules

For the systems that contain p shells (LiH, HF, N2, CO, H2O, CH4 and the free atoms C, O, F, Si, S, Cl with their variants) E_1 - Tr((H0 + H_ACP) P) is not zero. It is reproduced to the print precision by the on-site exchange correction of M 43 / S 154 to 159, written per spin with the dual density (P S + S P)/2, with one positive constant per element and shell pair (s-p, p-p) determined by least squares from the same data: 32 equations, 13 constants. With these constants and the polynomial radii measured from one molecule per element, every s/p reference system of the unmodified and the ACP-off parameter sets (48 runs, including NH3, H2CO, formamide, H3O+, OH-, OH and CH3) agrees to 4.3e-8 Eh. The constants are not stored; the test derives them each time (test_sp_systems_are_consistent_with_one_set_of_onsite_constants).

Electrostatics: first to fourth order

  • Preprint (M 25 to 31, M 36 to 40, S 82 to 106, S 129 to 144): shell charges population minus reference (S 99), third-order kernel (k3 U R - 1/2 k3x k3 U^2 R^2) exp(-k3x U R) with the mean Hubbard parameter U, atomic fourth order Gamma4/24 sum_A q_A^4.

  • Program: q_l = n^ref_l - population; f(q) = 1 + 0.012 [erf(q - 2/3) + erf(q + 2/3)]; mu_l = mu0_l (1 + k sqrt(CN)); off-site first order with Z_l - n^ref_l; U_l = u_l G_Z (1 + k sqrt(CN)); second-order kernel 1/(R + exp(-k2x R)/eta), eta the harmonic mean, R = 1e-12 on the atom; third order 1/3 sum q_i q_j q_A Gamma_A tau_ij with tau = -gamma^2/2 on the atom and (1.6 x - 0.92 x^2) exp(-1.15 x) between atoms, where x = R/eta (the inverse of the harmonic mean, not the mean U); fourth order shell-resolved, sum_l 0.021/4 q_l^4.

  • Evidence: the printed ES1 and ES2+3 of about 700 free-atom states and 360 molecules at the program’s density to 2e-8 and 7e-9 Eh (tests/test_gxtb_terms.py); the readings x = eta R and the preprint’s mean U miss H2+ and every heteronuclear molecule by 1e-3 to 0.5 Eh; ES1/f of all states of an element agree to 1e-8 only with the three constants of f.

Anisotropic electrostatics

  • Preprint (M 32, 33, S 108 to 117): one damping steepness per order, R0_AB the D3 pair radii, f_n tending to k_n.

  • Program: steepness 0.45 for R^-3 and 1.05 for R^-5, R^-7, R^-9; k9 = -0.003 (not in the file); R0_AB in angstrom against R in bohr; GFN2-xTB moments with the traceless quadrupole 3/2 rr - 1/2 r^2; only the q-d, q-Q and d-d parts enter the Fock matrix.

  • Evidence: printed ES multipole of every molecule to 5e-9 Eh.

Mulliken exchange and the on-site exchange correction

  • Preprint (M 41 to 43, S 146 to 159).

  • Program: per-spin form of S 152; fraction 0.15, same-shell factor 1.39, c_ij = sqrt(c_i c_j) from the two stored globals otherwise; U^x_l = u^x_l G_Z without CN; the switching exponent k1 + k2 R0_AB with R0_AB in angstrom; xi = 2 for polarisation shells: the d shells of the main-group elements and of Ca, the p shells of Li, Na, K, Zn and Ce, none for Be to Ne and the 3d metals.

  • Evidence: printed Ex (Mulliken) of every run to 5e-9 Eh; with the element hardness in double precision the free-atom ratio scatters by 2.8e-8, with it rounded to single precision by 3.8e-9. Every other flag pattern of an element misses its molecules by 1.7e-5 Eh or more.

  • The on-site correction (per-spin dual density, one constant per element and shell pair) reproduces E_1 of free atoms, h0_off, poly_off and acp_off runs (1078 runs) to 1.9e-7 Eh with the decoded H0 constants (2.8e-8 with them free), and E_1 of 501 unmodified and dispersion-free molecules, an independent set, to 4.2e-7 Eh (largest: SO2, CeF4, ArH+ triplet).

  • Fock matrix: the program enters the population part of the on-site correction twice off the diagonal, so its SCF solution is not stationary. gxtb.py reproduces this by default (fock_convention="program") and offers the exact derivative as the named alternative "variational".

Spin polarisation

  • Preprint (M 34, 35): W_ll’ = k^W_A W0_ll’.

  • Program: as printed; Espinpol of every open-shell run to 8e-9 Eh with the measured constants.

  • Measurement: a shell pair enters only through runs that magnetise both shells independently. The free atoms and ions of charge -1 to +3 never magnetise the d shell of Mg, Al, Si, Ga, Ge or the p shell of Ca, Sc, Ti (Aufbau fills d before p there), and magnetise the s shell of F and Kr only together with a full p set. The free ions of gxtb_v110_spin_ions.json (charges -7 to +7, every spin, 241 runs; 47 extreme anions whose molden.input or printed populations are incomplete are left out) and the low-symmetry mixing hydrides (_mix_) supply these equations. All pairs of all 38 measured elements are now determined (fit 6.0e-9 Eh over 757 runs, check 8.1e-9 Eh on the 11 remaining open-shell molecules). The new constants are short decimals (Mg dd 0.0213, Al sd -0.003225, Si dd -0.0067, F ss -0.2313), as the stored ones are; adding the rows moved a few earlier constants by up to 1e-5 (Si and Ge p-d, Ce s-s), which their earlier fits had absorbed into the then undetermined pairs; every printed Espinpol stays reproduced.

H0 constants for d and f shells

The third-party decoding gives K_d = 1.4, K_f = 2.85, F_d = F_f = 2 and k_delta = 5. A joint fit of the on-site exchange constants with these five constants and one polynomial radius per element free (free atoms and the h0_off, poly_off and acp_off runs of 360 molecules, among them TiH4, ScH, CuH, Cr2, Cu2, NiCO4, the 3d hydrides and CeF4, CeH4) gives K_d = 1.39999999 +- 5e-8, k_delta = 5.0000003 +- 6e-7, F_d = 2.0000004 +- 5e-7, K_f = 2.849996 +- 4e-7, F_f = 1.999996 +- 4e-6 (standard deviations from print-precision noise). All five are confirmed; the f values rest on cerium alone.

SCF

  • Program (decoded, checked): start density from n^ref and the EEQ_BC charges; damping 0.1 (single precision) in iterations 1 and 2 (open shell 1 to 3), a transition step, then Pulay mixing of 7 (8) Fock matrices; convergence at |dE| < 1e-8 Eh and rms dP < 2e-5 (open shell 4e-7 Eh, 5e-5) and one more unmixed iteration. vibe-qc’s driver gives the program’s iteration counts.

Open shells with a partly filled degenerate level

For about thirty open shells (3d hydride scans and probes, VH4, VO, CrO, CuCl4 2-, NO, Ti and Ce mixing hydrides, Ce triplets) vibe-qc’s total energy differs from the printed one by 1e-6 to 3e-4 Eh (CeH4 triplet 1e-3) and the iteration counts differ. Their highest occupied level is a partly filled, (near-)degenerate d, pi* or f set. What the program does there, with the evidence:

  • Occupations: T = 0 Aufbau, integer. molden.input of every such run lists occupations 0 and 1 only, and vibe-qc’s energy at the program’s printed density equals the printed total to 5e-8 Eh (the parity of the energy expression; VH4 4.6e-8, VO 3e-8). There is no fractional occupation, electronic temperature or level shift (the decoding agrees: lshift 0, no Fermi smearing at the default 0 K).

  • Same algorithm, same start. NiH (quartet, 1.40 x r_e): vibe-qc’s iteration energies equal the program’s printed ones to 1.5e-8 Eh for 15 iterations; from iteration 16 the paths separate (8.6e-7, 6.3e-6, …, 4.1e-4 Eh at iteration 22).

  • The printed endpoint is set by round-off. Changing vibe-qc’s start density by a relative 1e-12 moves its converged NiH energy by up to 3.9e-4 Eh and its iteration count from 54 to 142 to 249; rotating the molecule rigidly (the program convention is not rotation invariant at the 1e-7 Eh level) moves the program’s own printed energy by up to 1.75e-4 Eh (99 to 261 iterations against 155). The loose open-shell test (|dE| < 4e-7 Eh, rms dP < 5e-5) is met somewhere along a slow drift through a nearly flat direction (the orientation of the hole within the degenerate set), so the printed number is not a stationary point. It cannot be followed by any implementation that is not bit-identical.

  • The stationary state is the same in both programs. vibe-qc’s SCF converged tightly (|dE| < 1e-11 Eh, rms dP < 1e-8) from its own start and from the program’s printed density reaches the same state in 42 of 43 runs (to 2e-8 Eh). Written to gxtbrestart (packed spin densities and orbitals in the program’s AO order; p as x, y, z; d in the order x2-y2, z2, xy, xz, yz of vibe-qc’s real harmonics; f as vibe-qc’s), it is read by the program, which stays there: three iterations, a second restart moves it by less than 3e-9 Eh, and its energy equals vibe-qc’s to 5e-8 Eh or better, the energy-expression parity of these systems (gxtb_v110_stationary.json, generator gen_gxtb_stationary_states.py, test tests/test_gxtb_scf.py). Relative to the printed number, that state lies 8.5e-5 Eh below (VO) to 8.6e-7 Eh above (CoH probe at 1.54 angstrom, dispersion off); the program convention is not variational, so a loosely converged density can lie below or above it. The program’s own restart from its printed density drifts slowly towards the same state (NiH: -7.545e-5 Eh after four restarts, the state at -7.547e-5).

  • Two states. CrO with dispersion off has two stationary states 9.7e-7 Eh apart: vibe-qc’s start density reaches the lower one, the program’s printed density the upper one, in which the program also stays.

  • Not settled: CeF4 and CeH4 triplets. The tight SCF does not converge within 3000 iterations from either start (CeH4: -1.0e-3 Eh from vibe-qc’s start, -3.0e-3 Eh from the program’s density, both still drifting); the f manifold has several nearly degenerate states, and no stationary-state comparison is claimed for them (they are not in the stationary file).

vibe-qc keeps the program’s procedure and thresholds as the default (gxtb_single_point), so well-defined SCF solutions reproduce the printed numbers and iteration counts; for a degenerate open shell the reproducible quantity is the tightly converged state (energy_threshold, density_threshold; the analytic gradient driver converges tightly anyway).

Dispersion (revD4)

  • Preprint: M 44, S 161 to 175; D4 (Caldeweyher et al., J. Chem. Phys. 150, 154122 (2019), Eqs. 2 to 11) with the D4S weights (Tkachenko and Head-Gordon, J. Phys. Chem. Lett. 15, 10629 (2024)). The reference data are fetched at run time (gxtb_revd4_data.py, maintainer decision of 2026-11-14, see docs/license.md); only data are taken from that source, no code. The algorithm below is vibe-qc’s own (gxtb_dispersion.py), written from the papers and settled against the program.

  • What the program prints: dispersion is the two-body energy. The three-body (ATM) energy is in the total but not printed; it is recovered as total minus electronic, increments, dispersion and repulsion (noise about 2e-8 Eh). Neither is part of electronic.

  • Isolating variants of .gxtb (tests/data/gxtb_v110_dispersion.json, generator examples/regression/gxtb/gen_gxtb_dispersion_reference.py): a1_0 (no damping: the two-body term is -C6/R^6 - s8 C8/R^8, the three-body term undamped), s8_0 (no C8 term), disp_off (a1 = 1000, everything off).

  • Program reading, each item with its evidence:

    • Coordination number: the erf counting function of D4 with steepness 7.5, the pair radius the sum of the two revD4 covalent radii (4/3 of the Pyykko radii), no electronegativity factor. Evidence: the CN column of the program’s population table (two decimals) for every molecule, and the two-body parity below.

    • Reference weights: the D4S form with the element-pair factor of the revD4 pair table, so the weights of atom A depend on the element of its partner B; reference r of an element gets c(c + 1)/2 Gaussians, c the number of references of that element whose rounded counting CN equals that of r (the count of CN 0 starts at one, rounded CNs are capped at 19), as in the D4 reference implementation; when every Gaussian underflows the reference of the largest CN gets weight one. Evidence: with the D4 global factor 6 in place of the pair factor, benzene misses the printed two-body energy by 2e-4 Eh.

    • Reference polarisabilities: the reference polarisability of D4 with the hydrogen-type secondary system subtracted (D4 Eq. 5, scaled by the reference scaling factor, clipped at zero); the secondary system is scaled with the tanh function of its own element, T_S(q_h)/T_S(0), at the secondary-system charge q_h of the revD4 table, not with the D4 exponential charge scaling. Evidence: the exponential scaling misses LiH, formamide and TiCl4 by 1e-4 Eh, no scaling by 2e-4 to 2e-3 Eh. The secondary-system index is used as the atomic number of its element, as D4 does.

    • Charge scaling: zeta_r = T_A(q_A)/T_A(q_ref,r) with T = a + b tanh(c q

      • d) (S 167) and the self-consistent Mulliken charge q_A. Whether q_A is the valence charge minus the Mulliken population or the sum of the reference shell occupations minus it (they differ by up to 2.5e-7 e) is not resolved by the print precision: both give rms residuals of 2.8e-9 Eh over 273 molecules, the rms of pure rounding; vibe-qc uses the valence charge.

    • C6: Casimir-Polder integral by the 23-point trapezoidal rule of D4; C8 = 3 C6 Q_A Q_B with Q = sqrt(1/2 / sqrt(Z)).

    • Two-body damping: rational, R^n + (a1 R0)^n with R0 = sqrt(C8/C6) and no a2, s6 = 1, s8 and a1 from the file. Evidence: the a1_0 and s8_0 variants of Ar2, Ar3 and benzene to print precision.

    • Three-body term: s9 = 1.75, a program constant absent from the file (the a1_0 variant of benzene gives s9 = 1.75000 from a three-body energy of -1.1e-2 Eh; the benzene dimer and Ar3 agree). C9 = sqrt(|C6_AB C6_AC C6_BC|) with the C6 at zero charges, that is with zeta_r(0) = T_A(0)/T_A(q_ref,r), and the same pair-dependent weights; the SI’s “scaling function unity” misses ScF3, BF3 and CeF4 by 25 %. Zero damping 1/(1 + 6 (a1^3 R0_AB R0_AC R0_BC / (R_AB R_AC R_BC))^(16/3)) as S 171; positive for an equilateral triangle.

    • Cutoffs (sharp): no two-body term for pairs beyond 60 bohr (Kr2 at 59.9 and 60.1 bohr); no three-body term for a triple with any pair distance beyond 20 bohr (Kr3 with legs of 19.9 and 20.1 bohr; a bent Kr3 whose only long pair is 20.28 bohr has none).

    • Fock matrix: F = -1/2 (v_A + v_B) S with v_A = dE2/dq_A, the symmetric form of S 174, added to the other charge potentials; the three-body term has no Fock contribution. Evidence: total energies and SCF iteration counts below.

  • Derivatives for the analytic gradient: revd4_charge_potential (dE/dq_A) and revd4_nuclear_gradient (dE/dR at fixed charges, CN dependence of the weights included) agree with central differences to 1e-10 and 1e-9 (tests/test_gxtb_dispersion.py).

molden.input

  • f functions are written in the order xxx, yyy, zzz, xxy, xxz, xyy, yyz, xzz, yzz, xyz, all with the normalisation of xxx (d: each component normalised). For some high-spin free atoms the file lists fewer occupied orbitals than the program occupies (Ar and Kr anions with nine unpaired electrons); such runs are excluded by the electron-count check of gxtb_oracle.

Compiled-in tables: measured

Maintainer decision of 2026-11-14: measure what the program determines to print precision and store it with provenance (python/vibeqc/semiempirical/methods/gxtb_measured.py, protocols in gxtb_measure_protocols.py, regeneration examples/regression/gxtb/measure_gxtb_tables.py, test tests/test_gxtb_measured.py); fetch only what cannot be measured. The measurement runs are tests/data/gxtb_v110_measurement.json (same generator, --set measurement: free atoms and ions of 38 elements, H to Kr, I and Ce, in charges -1 to +3 and every spin; one hydride per element at several geometries with every variant; open-shell molecules; the reference molecules with dispersion off).

Table

Protocol

Result

element hardness G_Z

free-atom ES2+3 and on-site Ex

equals the D4 chemical hardness (dispersion_d4_model.zeta_hardness) rounded to single precision; referenced, not stored

reference occupations n^ref_l

ES1 + ES2+3 of all free-atom states, nonlinear least squares

38 elements, sd 2e-9 to 1.6e-7; fit 1.8e-8, molecules 2.1e-8 (ES1) and 6.5e-9 Eh (ES2+3); SI Tab. 1 agrees to its three truncated decimals; the sums deviate from the valence charge by up to 2.5e-7

third-order Gamma_A

the same fit

two-decimal values (raw within 3 sd, sd <= 4.4e-6), all negative

valence shell charges Z_l

ES1 of molecules

Aufbau: s first, 3d metals s2 d(n-2), Ce s2 d1 f1

polarisation shells

Ex of molecules, every flag pattern

see Mulliken exchange above

spin constants W_ll’

Espinpol of free atoms and ions (charges -7 to +7) and open-shell molecules, linear

fit 6.0e-9, check 8.1e-9 Eh; every pair of every measured element determined

on-site exchange constants

joint fit with H0 (above)

151 constants, sd 1e-9 to 1.2e-4

H0 polynomial radii

joint fit with H0

the DFT-D3 covalent radii (gxtb_aux_data, Apache-2.0) to <= 1.6e-6 angstrom for every element but H (0.418 angstrom) and F (0.500 angstrom); referenced, only H and F stored

revD4 reference data

not measurable

fetched at run time from axtb (maintainer decision 2026-11-14, gxtb_revd4_data.py); s9 = 1.75 and the 60 and 20 bohr cutoffs measured (see Dispersion)

An element or shell pair without a measured value makes the calculation refuse (GxtbMeasuredTableError, or a missing-constant error of the term).

Parity of the numpy implementation

vibe-qc’s own SCF (gxtb.py, program Fock convention, its own start density) against the program’s 14-digit total.

With the revD4 dispersion, every converged base run of the reference, measurement and dispersion sets (999 runs, measured 2026-11-14; the 129 open-shell runs that were refused until all spin constants were measured are included):

Class

Runs

max abs dE(total), Eh

Same iteration count

neutral free atoms (no dispersion)

121

8.6e-9

all

free ions

364

1.4e-8

all

closed-shell s/p molecules (H to Ne), benzene dimers and hexane included

78

4.7e-8 (BF3)

all

closed-shell molecular ions (H to Ne)

15

4.7e-8 (BF4-)

all

closed-shell heavier main group (Na to Kr, I), noble-gas clusters and cutoff probes included

141

3.0e-7 (SO2; SF6 1.7e-7, I2 dimer 1.1e-7)

all

closed-shell d and f molecules, ferrocene included

39

2.2e-7 (CeF4)

all

open shells within 1e-7 Eh

68

6.7e-8 (HCl triplet)

65 of 68

open shells 1e-7 to 1e-6 Eh

14

9.7e-7 (VH at 1.40 x r_e)

5 of 14

open shells beyond (degenerate: V, Cr, Fe, Co, Ni, Zn hydride scans and probes, Ti and Ce mixing hydrides, VH4, CuCl4 2-, CrO, Ce triplets)

30

1.0e-3 (CeH4 triplet; else 6.1e-5, CeF4 triplet)

1 of 30

The closed-shell deviations equal those of the same runs without dispersion. The open shells beyond 1e-6 Eh, and some between 1e-7 and 1e-6 Eh (NO, the CrH and NiH scans), are the degenerate cases of “Open shells with a partly filled degenerate level” above: their printed numbers are loosely converged points of a drifting SCF, and which of them land within 1e-6 Eh is chance (refitting the spin constants, which moves the energy at a fixed density by about 1e-9 Eh, reorders them). Their stationary states agree with the program’s to 5e-8 Eh (42 of 43; CrO without dispersion has two states; the Ce triplets do not converge tightly). The printed two-body energy at the program’s density agrees to the print precision (rms 2.8e-9 Eh over 273 closed-shell molecules, the rms of rounding), the three-body energy to its noise (1.2e-8 Eh in the dispersion set), and the dispersion contribution E(base) - E(disp_off) to the 14-digit total of the 17 systems of the dispersion set that carry both variants (benzene dimers, hexane, ferrocene, I2 dimer, noble-gas clusters, cutoff probes, open-shell Ar complexes) agrees to 3.3e-10 Eh (tests/test_gxtb_dispersion.py).

Without dispersion (free atoms and ions; molecules of the disp_off variant), measured 2026-11-14:

Class

Runs

max abs dE(total), Eh

Same iteration count

neutral free atoms

87

8.6e-9

all

free ions

292

1.4e-8

all

s/p molecules (H to Ne)

28

4.7e-8

all

heavier main group (Na to Kr, I)

50

3.0e-7 (SO2)

all

closed-shell d and f molecules

14

2.2e-7 (CeF4)

all

open-shell molecules within 1e-7 Eh

33

6.7e-8 (HCl triplet)

30 of 33

open-shell molecules 1e-7 to 1e-6 Eh (HI and ArH+ triplets, VH2, VH, CeO, Cl2+, Br2+, I2+)

8

8.5e-7 (Br2+)

4 of 8

open-shell molecules beyond (degenerate: NO, CuCl4 2-, VH4, VO, VH triplet, Cr, Co, Fe, Ni hydrides, CrO, Ce triplets)

14

1.0e-3 (CeH4 triplet; else 8.3e-5)

0 of 14

The deviations above 1e-8 Eh of well-defined solutions follow the least-determined on-site exchange constants (the E_1 check of the measured tables is 4.2e-7 Eh). Where a degenerate HOMO set is only partly filled the printed endpoint is set by round-off (above; the third-party decoding, Sec. 4.17, reaches the same conclusion), and the comparison is between stationary states. No open-shell run is refused for a missing spin constant any more (the free atoms and ions add 21 neutral and 68 charged open-shell states, all within 8.4e-9 Eh with the program’s iteration counts).