Liquid water is proposed to be a fluctuating mixture of two local structures: LDL (low-density liquid — tetrahedral, open, 4-coordinated, ice-like short-range order) and HDL (high-density liquid — disordered, collapsed, an interstitial 5th neighbour pushed in). Note both are liquid: the split is within the liquid, not ice-versus-water. The open field question underneath the debate is what molecular feature makes the tetrahedral LDL structure form at all.
| Experiment | Result | Reading |
|---|---|---|
| Dimer, reduced O (Z=2, isotropic) | O–O holds ~3.0 Å — stable H-bond | the isotropic cloud binds |
| Liquid, reduced O | disordered, coordination ~2.6, no LDL | isotropic acceptor ⇒ no tetrahedral preference |
| Dimer, carved lone pairs (3-split +3; hemisphere split +2) | O–O grows — dissociates | fragmenting the cloud destroys the H-bond |
| Ice (diamond lattice, reduced O) | q ≈ 1, coordination ≈ 4 — holds the tetrahedron | directional donors alone keep an imposed lattice |
| Melting ice (fixed volume) | q 0.85 → 0.55, coordination stays ~3.3 | heat disorders but can't densify ⇒ not HDL |
| Compression (density ↑) | forces a 5th neighbour in — coordination climbs, d5 collapses | HDL is a density effect, needs no directionality |
The natural fix — give oxygen directional lone-pair lobes — was tried three ways (a 3-fold split, and a charge-neutral hemisphere split biased toward the acceptor side). Every one dissociated the dimer, while the plain isotropic water stayed bound. The reason is mechanistic: in RealQM the H-bond is the donor proton binding to a single coherent acceptor cloud. Splitting oxygen into angular domains puts a free boundary through that cloud and breaks the coherent density the proton was holding onto. You cannot have directional acceptors and binding at once at this reduction. (The angular-split machinery itself works — validated on a single molecule — it is the water physics that resists.)
Textbook water is tetrahedral because oxygen has 2 directional donors (O–H bonds) and 2 directional acceptors (lone pairs). RealQM's reduced water has the donors but an isotropic acceptor. That single difference reproduces a coherent story:
So the parameter-free experiment points a finger at the driver: spontaneous LDL requires directional lone-pair acceptors; donor directionality is not sufficient. And it reframes the pair cleanly — HDL is the “default” isotropic-packing structure; LDL is the special, acceptor-stabilised one. That is a falsifiable, parameter-free statement about the origin of the two structures — the mechanistic question underneath the “does water have two liquids” headline.
Viscosity is an independent, transport-level test — and it ties straight to the H-bond picture, because in water viscous flow is H-bonds breaking and re-forming as layers slide (so the activation energy for flow ≈ the H-bond energy). We drive a Couette-like shear (a body force Fx = mγ(y−yc)) and read η = stress / strain-rate ≈ γ / (dvx/dy) from the velocity profile.
The catch — and how we beat it. Under a strong local thermostat, F/v measures the thermostat friction Γ, not the fluid (the profile goes linear, not parabolic). We separate the two by how each responds to the thermostat: weakening it 4× (γL 0.02→0.005) dropped the high-T plateau from ~0.009 to ~0.005 (it scales with Γ — that's the thermostat floor), while the cold value barely moved (0.015→0.012 — that's the fluid). The spread across T widened (1.7× → 2.4×): lowering the floor let the fluid thinning show.
Result: the extracted fluid viscosity (~0.007 model units cold, thinning on heating) has the correct sign (water thins when heated) but a shallow slope — a small activation energy, exactly the weak, under-coordinated H-bond signature. Honest limits: model units (no Pa·s — no calibrated time), and the temperature scale is itself uncalibrated (melts ~1300 “model-K”), so only the sign, relative thinning, and the thermostat decomposition are meaningful. Within that: a parameter-free model reproduces water's qualitative viscosity and separates fluid response from the thermostat friction floor. ▶ shear-viscosity probe
Why is the viscosity small — and the isotropy coin again. By the Maxwell relation η ≈ G∞·τ (stiffness × relaxation time), a viscosity is small either because the network is soft (small G∞) or because it relaxes fast (small τ). Water is the fast kind — stiff yet fluid, because H-bonds re-route by low-barrier concerted switching, cheap precisely because the acceptor is angularly soft. So the same isotropy that denies LDL (no tetrahedral preference) also enables facile switching (low viscosity): no-LDL and low-viscosity are dual consequences of one property. Honest caveat: extracting G∞ and τ under a driven shear was unreliable — a weak-enough thermostat let the shear heat the system into runaway (the classic NEMD thermostat dilemma), so a clean split needs the equilibrium Green–Kubo route (stress autocorrelation, no driving). The qualitative result (correct-sign, shallow thinning) is unaffected.
We tried to pin down precise coefficients too, and record it honestly: the shear viscosity was only qualitatively accessible (a weak-thermostat driven shear ran away in temperature — the NEMD thermostat dilemma), and a virial bulk modulus was dominated by intramolecular stress (so the intermolecular liquid modulus couldn't be cleanly separated). Clean values need equilibrium Green–Kubo (η) and an energy–volume modulus K=V·d²E/dV² (compressibility).
But both landed in the correct regime, parameter-free: water is runny (small viscosity) and stiff (small compressibility) — from complementary origins, a soft acceptor (fast H-bond switching → low η) and hard electron exclusion (→ low compressibility). And this is arguably what matters: macroscopically water is an essentially incompressible, nearly inviscid fluid (low Mach, high Reynolds), whose large-scale flow depends on these being small far more than on their exact values. So the precise coefficients being elusive here is consistent with their macroscale near-irrelevance — what counts is that a parameter-free model places water in the right small-η / small-compressibility regime, which it does.
Fair question — and it splits in two, with opposite answers.
So the framing is complementary, not competitive: DFT would be more accurate for real water's structure; it can't cheaply reach the supercooled two-state regime; and it structurally can't run the ablation. The reduced RealQM water isn't competing on accuracy — it's a conceptual microscope that isolates mechanism at laptop cost.
We never saw a clean bimodal q at one temperature (LDL and HDL coexisting). That is not a failure of the runs — it is the result: the reduced model cannot produce the coexistence, because it lacks the directional acceptor that would make tetrahedral order compete with dense packing. A model that both H-bonds and carries a directional acceptor — which domain-splitting cannot deliver here — would be needed to test the two-state model itself. That is the open problem this work isolates.
The crisp “two equivalent tetrahedral lone pairs” picture is a localized-orbital model, not a unique observable: photoelectron spectra show water's two highest orbitals sit at different energies (not equivalent twins), and the electron density shows a smeared acceptor-rich crescent, not two sharp lobes. What is robustly observed is the 4-coordination and an anisotropic, acceptor-rich oxygen density — so a faithful model needs directional acceptance, not literal “rabbit-ears.”
Interactive runs: reduced-water probe · melt-ice two-structure probe · compression → HDL scan · dimer control · 3-split water (port test)