Forces at the observed geometry

A test of a molecular model that needs no reference calculation, no energy scan, and no minimum-finding

The idea. At a molecule's measured geometry the true force on every nucleus is zero by definition. So compute the forces there: whatever remains is model error, read directly. No reference calculation is needed, because experiment supplies the answer; no minimum has to be located, because the geometry is given; and one calculation per molecule replaces a scan over hundreds.

The residual converts straight into the quantity of interest. A force F at the observed geometry displaces the bond by ΔR ≈ F/k with the measured force constant, so "the forces are small" becomes a bond-length error without ever computing a bond length. For H2, k = 5.16 mdyn/Å = 0.3314 Ha/a02, so F = 0.01 Ha/a0 means the model's bond is off by 3%.

This matters for RealQM specifically because the model is a force law — kernels move under the Coulomb force of the charge distribution — so the force is primary and the geometry is wherever it vanishes. The method tests the model on its own terms.

First the force had to be fixed

The solver computed the force as Z·∇VH at the nucleus, finite-differenced and averaged over a sphere of FORCE_RADIUS cells. Holding H2 at a fixed R = 1.68 and varying only that radius:

FORCE_RADIUS (cells)0.24 a₀0.48 a₀0.72 a₀0.96 a₀
F−0.05019−0.03227−0.00624+0.02706

No plateau, and the sign flips. The quantity being reported was the smoothing radius, not a force. Two causes: ∇VH is unresolved at the nuclear cusp on this mesh, and the averaging bias is O(a²∇ρ) with ∇ρ largest exactly there.

The replacement is the analytic derivative of the discretised energy. Since VeK = −Σa Za∫ρ/rsoft with rsoft2 = |r−Ra|² + (SOFT·h)², differentiating gives

Fa = Za ∫ ρ(r) (r − Ra) / rsoft3 dV

a volume integral: no averaging radius exists to tune, nothing is evaluated at the cusp, and because it differentiates the same softened kernel the energy is built from, agreement with −dE/dR holds by construction. Validated on H2, with force and energy from the same four runs: the volume force crosses zero at 1.84–1.89 against an energy minimum at 1.963 — 4–6% — where the point gradient disagreed by 0.42 a0 and had the wrong sign at R = 1.68.

What can be checked — and why it depends on symmetry

Three independent checks ride on every run. Which of them carries information is decided by the molecule's symmetry group, and conflating them is easy:

VKK — exact arithmetic

The kernel–kernel energy is Σa<b ZaZb/Rab, computable by hand. It confirms the geometry actually built — every row below matched its predicted value, which is how fractional kernel placement and per-bond spacing were verified.

ΣF = 0 — a convergence meter, not an identity

The total force on the nuclei is not automatically zero: it equals the net force the electron cloud exerts on them. Only when the electronic state is converged is the cloud stationary, and then by Newton's third law the sum vanishes. So ΣF measures electronic non-convergence — and it does so in a way no energy-based criterion here could, since the solver's own convergence flag compares energies between wall-clock polls and reports CONV on states 30 mHa from settled.

But symmetry cancels it. In a symmetric molecule the mirror or rotational symmetry forces ΣF = 0 whatever the accuracy. CO2's ΣF = 0.0001 therefore proves only that the algorithm preserves symmetry — it says nothing about the forces being right. ΣF informs only in directions symmetry does not protect: all directions for a heteronuclear molecule, the C3 axis of NH3, the C2 axis of water.

Per-atom zeros — what symmetry cannot fake

An atom sitting at a centre of symmetry must feel exactly zero: the central C of CO2, the C of CH4. Unlike ΣF this is a single atom's force, so a nonzero value is unambiguously error and its size is the per-atom floor.

Results

All at measured geometries, bare kernels (rc = 0), convergent functional, 100³ grid, box 12 a0, 120k steps. + = outward (bond wants to lengthen).

moleculeZxforceΔR via F/kΣF / |F|maxstatus
H21,1+0.1383+30%14.7%marginal
N23,3−0.2505−8.2%3.7%good
O22,2−0.1682−9.7%11.0%fair
O23,3−0.2463−14.3%8.4%fair
CO4,2+0.304 / −0.187—38.5%fails
CO2,2−0.176—13.2%degenerate
CO22,4,2−0.1632−7.2%0.06%*good
CO22,2,2−0.1575−7.0%0.02%*degenerate
HCN1,4,3C≡N −0.193, H−C +0.026−7.8% / +3.5%5.5%good
HCN1,2,3−0.152—88.7%fails
H2O2,1,1O −0.197 (y)—140%fails
H2O3,1,1O −0.393 (y)—126%fails
H3 linear1,1,1+0.0420—34.9%fails
H4 chain1,1,1,1+0.0316 outer— (see FCI)7.4–16%partial
C2H21,4,4,1all outward—25%fails
C2H21,2,2,1all outward—28%fails
C2N23,4,4,3N in, C out—3.2%good
NH33,1,1,1net +z 0.382—141% (z)fails
NH3 planar3,1,1,1——70% (x)fails
CH44,1,1,1,1H 0.036 outward—0.6%best
CH42,1,1,1,1——40%fails

* symmetry-enforced — carries no accuracy information. "degenerate" means the Zx choice makes two different elements identical, so the molecule is not the molecule named.

H4 against an exact reference

The one row scored against an exact method rather than experiment. FCI (sto-6g) puts the equally spaced chain's equilibrium at a = 1.8 a0; one electron per site, bare kernels, no rc to argue about. At that geometry all four RealQM forces point outward — the model wants to expand where FCI says equilibrium — while over-binding by 14–18% per atom (2.37–2.45 eV against FCI's 2.083, each referred to its own atom). Over-bound and too long is the same pattern H2 shows at +30%.

What the method revealed

0. KERNEL DEPTH decides convergence — two controlled comparisons

The sharpest results of the study hold geometry, symmetry and every other input fixed and change only the central kernel's charge:

moleculeshallowdeepfactor
CH4 (identical Td geometry)C+2: 40%C+4: 0.6%60×
HCN (identical geometry)C+2: 88.7%C+4: 5.5%16×

A deep kernel holds its domain; a shallow one does not, and the boundary against its neighbours stays undetermined. Note the direction: C+4 is the less resolved kernel (Z·h = 0.48 against the Z·h ≲ 0.1 guideline), so depth helps convergence more than coarse resolution hurts it.

Three alternative explanations were tested and excluded.

0b. A second, smaller effect: cubic-grid anisotropy

Planar NH3 (D3h) must have ΣF = 0 in every direction and gives ΣFx = 70%. Its three ligands: the one lying on the grid x-axis has radial force 0.0802, the two off-axis ones 0.0545 and 0.0544 — agreeing with each other to 0.2% and differing from the first by 47%. A cubic grid has 4-fold symmetry in-plane and 3-fold about its body diagonals, so molecular 3-fold and bent geometries cannot be represented evenly, while linear and tetrahedral ones can. That is why NH3 and H2O are the two worst rows and CH4 is among the best.

1. Convergence, not chemistry, is the blocker

Ranked by ΣF in unprotected directions, the failures are not where "harder chemistry" would put them. They track how tightly the electron domains are held: N2 (Z=3) 3.7%, HCN with C+4 5.5%, H2 14.7%, linear H3 34.9%, HCN with C+2 88.7%, water 140%. Deep kernels settle; diffuse hydrogen-bearing systems do not, even at 120,000 steps.

This inverts an expectation worth recording: C+4 converges far better than C+2 in HCN (5.5% vs 88.7%) despite being the less resolved kernel (Z·h = 0.48 against the Z·h ≲ 0.1 guideline). Depth helps convergence more than coarse resolution hurts it.

2. Interior domains are the ones that fail

The H4 step ladder separates them. Between 60k and 180k steps the outer atoms' forces move 2% while the inner pair moves 20–30% and their mutual asymmetry grows from 1.4% to 50% — worse with more iteration, not better. So the failure is not "more atoms" and not "three centres": it is domains with neighbours on both sides.

3. The recorded three-centre failure looks like an artefact

Every three-centre system previously tried "descended monotonically inward through its experimental bond". Measured by force instead, CO2 gives −7.2% and HCN −7.8% — ordinary bounded model errors, not a collapse. Those scans sampled an oscillating relaxation (linear H3 swings ~0.03 Ha out to 300,000 steps) through a convergence flag that could not detect it. They need rerunning, not explaining.

4. Zx is testable — and the house conventions mostly survive

A depth ladder at one mesh (box 16, h = 0.145), scored by pair asymmetry and by bond error:

O+2O+3O+6 (one blob)N+3N+5
pair asymmetry19.5%3.7%1.4%6.0%6.1%
ΔR−6.4%−7.1%+59%−8.1%−13.7%

O+3 is the best oxygen reading — a 5× convergence gain over O+2 at the same bond error. N+3 stands: full valence costs accuracy and buys no convergence. And O+6 as a single blob converges best while giving a +59% bond — converging beautifully to the wrong answer, which independently confirms the atomic work's rejection of the [2,6] configuration. So "use the deepest Zx" is false past the valence structure the atom model selects; depth helps only where the domain was genuinely under-bound.

It is notable that O+3 is exactly one of the two valence shells in the atom model's 2+3+3 — atomic and molecular evidence converging on the same number independently. The faithful 3+3 construction was attempted here as two coincident +3 kernels; the geometry built correctly (VKK exact) but the two domains superposed instead of nesting (Vee = 72 Ha, E positive), because identical seeds in an identical environment have nothing to break the degeneracy. Doing it properly needs the shell machinery — one-electron domains with per-shell z_init, as mol_fast_atom_split.html does — and 6 domains per oxygen exceeds this page's 8-domain cap.

4b. Two readings are degenerate

For O2, O at +2 gives the smaller residual than +3 (−9.7% vs −14.3%). But Zx is the only thing distinguishing elements in this construction, so C+2/O+2 makes carbon and oxygen the same atom: "CO at 2,2" is a homonuclear X2 and "CO2 at 2,2,2" is X3. Those rows carry no information about the named molecule.

Scope. Bare kernels throughout, so Z·h = 0.48 for C+4 — five times the resolution guideline — and no ΔR here is quotable without a mesh ladder. The directions, the error bars, and the convergence ranking do not depend on that. Rows marked "fails" exceed their own error check and are reported as measurements of the floor, not as results. Measured geometries, force constants and the FCI reference are literature values, flagged unverified in the source.

Why this is the right instrument for this model

Two routes to a computed geometry exist. Scanning E(R) needs every point converged, which at three centres they are not. Relaxation needs a trustworthy force, which until now did not exist. Evaluating the force at a given geometry needs neither: one point, one convergence question, and an answer supplied by experiment.

It also degrades gracefully. When a row fails, it fails visibly — ΣF exceeds the signal — rather than returning a plausible wrong bond length. Of the fourteen rows above, six announce their own unreliability. A scan would have returned numbers for all fourteen.

Code: mol_fast.js (USER_FORCE_VOLUME, USER_FORCE_RADIUS), mol_fast_h2_scan.html (?forcevol, ?frad, ?z= comma list, ?sep= per-bond list, ?geom=iso|cen, per-atom force and ΣF readout). Energies in Hartree, distances in bohr, forces in Ha/a0.