The two-centre test

Where the pseudo-atom core model works, and where it stops — measured on X2

RealQM replaces an atom's filled core by a pseudo-kernel: a reduced charge Zx carrying only the bonding valence, with the core represented by one number, rc. Everything downstream — molecules, crystals, proteins — rests on that substitution. This page reports what the substitution actually delivers, measured on the one class where the model produces a computed bond length.

The result in one line. The pseudo-atom works in a shallow, monovalent corner — alkali dimers, bonds to a few percent — and fails in three independent ways outside it. No choice of rc repairs this, because the bond error changes sign with core depth.

Why X2

Diatomics are the only systems here with a computed geometry: one scan coordinate, an interior minimum that can be bracketed, and E(X) available on the same mesh so the discretisation cancels in Ebind = 2E(X) − E(X2). Three-centre systems fail separately and for reasons not yet understood — H3 linear, O3, ClO2 and CO2 all descend monotonically through their experimental bond, across Z = 1…4, with and without a core, odd and closed-shell. Water's angle survives only because its bond is supplied as an input. So X2 is where the core model can be tested, and it is where every statement below was measured.

Three core models, and what rc actually is

modelconstruction
hard depletiondensity ≡ 0 for r < R, free (Neumann) surface at R
clampV = −Z/max(r, rc) — a plateau; density admitted inside
empty coreV = 0 inside rc

rc is not a radius. One knob sets a plateau whose depth is Zx/rc, and the ionisation energy pins the depth; the extent then merely follows as Zx/depth. Carbon's rc is twice oxygen's because its kernel charge is twice as large, not because its core is bigger. Nor is it an element constant — it is a constant of (element, bonding reading): O at +2 and O at +3 are different pseudo-atoms with different radii. And no conversion carries between the models: the hard/clamp ratio runs 0.66 at Zx = 1, 0.70 at 2, 0.78–0.81 at 3, 0.88 at 4.

The calibration

rc is fixed per element by requiring the pseudo-atom to reproduce that element's measured first ionisation energy, IE₁ = E(Zx−1) − E(Zx), computed in the same construction on the same mesh so the discretisation largely cancels, then bisected on rc. Zx electrons share one spherical density, pairs counted n(n−1)/2 by the i ≠ j rule — no (Z−1)/Z factor and no kinetic fudge. The radial code verifies E(H) = −0.49999583 against −0.5 and the hydrogenic Hartree integral 0.62500417 against 5Z/8, both to 4.2e−06; radii are converged to 5.7e−06 and agree with an independent 2D axisymmetric fit to ≤0.2%.

The measured IE is the model's one empirical input. That is the cost: the atom is no longer a prediction, so only bonds and binding remain as tests.

1. The shallow corner — where it works

Alkali dimers, each rc fitted to its own element's IE, same mesh for atom and dimer. Plateau depth Zx/rc = 0.21–0.33 Ha throughout.

modelmean |bond| errbinding err (Li, Na, K, Rb, Cs)
clamp4.0%−3, +22, +10, +11, +1%
hard depletion8.3%+43, +87, +98, +102, +98%
empty core11.4%−47, −52, −53, −64, −68%

This is the result the method is usually quoted on, and it is real: swapping only the core model, at the same calibration, takes binding from +86% to +10% and bonds from 8.3% to 4.0%. The penetration of valence into the clamped core (Cs 51%, Cl 26%, O 16%) is a feature of the working model, not a defect.

2. Depth — where it inverts

The alkali plateaus are all shallow. To vary depth alone the halogens are the right probe: they are singly bonded, so Zx = 1 is the complete bonding valence with nothing truncated, yet the plateau is deep because the atomic IE is large. Same monovalent code, same calibration, real molecules.

system1/rc (Ha)cells per rcDminRe expbond errbinding err
K20.24247.5257.379+2.0%+17.0%
Na20.31196.1255.818+5.3%+24.9%
Li20.3317 → 355.6885.051+12.6%−3.2%
I21.294.5 → 8.82.1885.038−56.6%+166.7%
Cl24.191.6 → 2.71.4873.756−60.4%+167.1%

Three features make this a verdict rather than a bad run:

Mechanism. The clamp admits density inside rc, so two equivalent domains facing each other across the midplane meet no Pauli repulsion between their cores. Nothing resists approach but VKK = Z²/D, which arrests the contraction only at ≈45% of the true bond. Shallow plateaus are too weak for this to dominate; deep ones are not. The same mechanism is on record for RealQM crystals, which all contract.

3. The second row — where both models fail

Zx > 1, IE-calibrated radii, one axisymmetric code, commensurate scans.

systemmodelrc / RDminRe expbondEbDe expbinding
O2hard0.547111.9252.282−15.6%16.44 eV5.12+221%
O2clamp0.786831.9252.282−15.6%12.14 eV5.12+137%
N2hard0.72670no bond — repulsive to 10.7 a₀—unbound9.76—
N2clamp0.937414.0252.074+94.1%4.86 eV9.76−50%
C2hard1.33246geometrically impossible——6.21—
C2clamp1.506486.825*2.348+190.7%*unbound6.21—

* minimum at a scan edge — not an interior minimum, and not to be read as a bond.

O2 lands on the same 15.6%-short bond in both models and over-binds by 2.4–3.2×. N2 is either unbound entirely or bonded at nearly twice the right length. C2 cannot be set up at all in hard depletion, and is unbound in the clamp.

Two structural results

A ceiling of one Rydberg

A monovalent clamped kernel cannot reproduce an ionisation energy above 13.6057 eV — its rc → 0 limit is hydrogen. Fluorine (17.42 eV) has no solution at any rc; neither would nitrogen (14.53) or oxygen (13.618, over by 0.012). So Zx > 1 for the second row is forced by the construction, not a house convention — and F2 cannot be built monovalently at all.

Calibrated cores do not fit organic bonds

Hard depletion forbids core interpenetration, so a bond requires RA + RB < d. At the IE-calibrated radii:

bondd (Å)d (a₀)ΣR calibrated
C–C single1.542.9102.665fits, margin 0.245 ≈ 2h
C=C aromatic1.392.6272.665impossible
C=C alkene1.342.5322.665impossible
C–N peptide1.332.5132.059fits
C=O carbonyl1.232.3241.880fits

Aromatic and double-bonded carbon cannot be represented at all — Phe, Tyr, Trp, His, every C=C. This is arithmetic, prior to any energy calculation. It also shows what the calibration is: the IE sets a depth that scales with Zx, and hard depletion then reads that number as an excluded volume. The IE calibration and hard depletion are mutually incompatible for carbon.

Consequence for protein-scale work. The protein pages run C 0.3, N 0.3, O 0.4–0.8 — four-fold smaller than the calibrated radii, and calibrated to nothing. They work because of that: the calibrated values would make the bonds geometrically impossible. Recalibrating the protein stack is therefore not an improvement to be applied on principle, and the existing results stand on their own validation against PySCF, not on the core model's provenance.

What this means

The honest scope of the pseudo-atom route, measured rather than asserted:

The open problem is sharply stated: a core model for the second row. Hard depletion treats rc as excluded volume and breaks on geometry; the clamp treats it as a penetrable plateau and over-binds by ≈2.7× because equivalent cores feel no exclusion. Neither interpretation of one number can be right for both, which suggests the single parameter is the problem rather than its value.

Method notes

Code

All offline, no basis set, no fitted parameter beyond the IE. Each occupancy generalisation was validated against its monovalent parent before use.

Measured values (IE, Re, De) are literature references used as targets and are flagged unverified in the source files; they are the one external input. Energies are in Hartree, distances in bohr unless stated.