Why ice floats
Nearly every liquid shrinks as it cools and again when it freezes. Water does the opposite, twice — because of the shape of one bond.
Two things water does wrong
Cool water and it gets denser — but only down to about 4 °C. Below that it expands again. At 0 °C it freezes and expands further. The solid is lighter than the liquid, so ice floats.
Most substances do neither: their solid is their densest form, and sinks.
Why it matters
Because ice floats, a lake freezes from the top down. The densest water, near 4 °C, sits at the bottom, and the ice lid insulates everything below. If ice sank, a cold winter would freeze a lake solid. The density maximum is also a standing test for any theory of liquids: one that cannot produce it is missing something about water.
Interactive Drag the temperature between −10 and 40 °C and watch the column height; untick hydrogen bonds to see the dip near 4 °C disappear.
The bond that makes room
A water molecule is one oxygen with two hydrogens. The oxygen pulls electrons toward itself, so it is slightly negative and the hydrogens slightly positive. A hydrogen bond is the attraction between a hydrogen on one molecule and the oxygen on a neighbour. It is much weaker than the bond inside a molecule, but it points in a definite direction.
Each molecule can make four: two through its hydrogens, two through its oxygen. The four neighbours sit at the corners of a tetrahedron — a pyramid with four triangular faces. Neighbours held at fixed angles leave gaps a loosely packed liquid would fill.
Bernal and Fowler proposed the consequence in 1933. At low temperature the network is nearly complete and ordered, like ice: large open volumes, low density. Heat it and the network loosens: bonds break and re-form every trillionth of a second, molecules slip into the gaps, and the liquid packs more tightly.
Two effects pulling in opposite directions
Heating water does two things at once. It loosens the network, letting molecules from the second ring of neighbours slip into the gaps: denser. It also makes every molecule jiggle harder, pushing neighbours apart: less dense. The second effect is ordinary thermal expansion; every liquid has it.
Just above freezing, much of the open second ring is still standing, and each degree of warming collapses a good share, so collapse wins and water gets denser. Higher up, the second ring has collapsed about as far as it can; the nearest four neighbours still sit close to their tetrahedron, but little open room is left to lose, so jiggling wins and water expands. The crossover is the density maximum, measured at 3.98 °C and 1.00 g/cm³.
Freezing is the extreme case. In ice every molecule has all four bonds in the full open network — the roomiest arrangement of all, so the solid takes more space than the liquid. This is the negative volume of melting: ice shrinks when it melts.
The modern picture sharpens Bernal and Fowler's idea. Simulations describe liquid water as a shifting mixture of open tetrahedral neighbourhoods and denser disordered ones, and a single number — how much order survives in the second ring — tracks the density anomaly well. Recent machine-learned simulations pin down where: the nearest four neighbours stay close to an ideal tetrahedron even at room temperature; the second ring is what collapses. Strong directional bonds set the shape; weak, non-directional attractions decide how far the gaps get filled.
In short
Water's hydrogen bonds hold its molecules in a roomy four-cornered arrangement. Heat lets molecules slip into the gaps (denser) and makes them jiggle (less dense). Below 4 °C the collapsing network wins, above it the jiggling wins. Freezing completes the network, so ice is roomiest of all and floats.
Where this comes from
- Understanding the Density Maximum of Water with Machine Learned Potentials linked only, not reproduced
arxiv.org/abs/2603.27767 - How van der Waals interactions determine the unique properties of water linked only, not reproduced
arxiv.org/abs/1606.07775 - Understanding water's anomalies with locally favored structures linked only, not reproduced
arxiv.org/abs/1308.4231