Liquid water reaches its maximum density at exactly 3.98°C, then expands as it continues to cool toward freezing. Every first-year chemistry student learns this fact. Almost none of them are told why it happens – and until very recently, neither could the scientists who studied water for a living.
That gap has never been from lack of effort. Researchers have cataloged more than 70 documented ways that liquid water defies the behavior expected of a simple liquid, from the way it compresses more easily as it cools, to the speed at which sound travels through it, to the way it behaves near surfaces at the molecular scale. Each anomaly, taken individually, is interesting. Taken together, they’ve pointed generations of physicists toward a single uncomfortable hypothesis: that what we call “liquid water” is not actually one thing.
A study published in June 2026 in Nature Physics has now provided the strongest molecular-level evidence to date that this hypothesis is correct. The water in your glass, the research suggests, is a constantly shifting mixture of two structurally distinct microscopic forms – and the interplay between them may be responsible for most of water’s strange and life-sustaining behavior.
Executive Summary
A study published in Nature Physics provides new molecular-level evidence from simulations that liquid water is not a single uniform substance, but a constantly shifting mixture of two distinct microscopic structures. The research, led by Prof. Xiao Cheng Zeng of the City University of Hong Kong, deployed unsupervised deep learning on an unprecedented dataset of simulated water molecules to identify two interconvertible local structures – one denser and disordered, the other lighter and more ordered. The water forms discovery settles a decades-long theoretical debate, offers a unified explanation for water’s most puzzling physical properties, and opens a new experimental frontier in understanding one of the most important substances on Earth.
A Century-Old Debate, Finally Addressed
The idea that water might exist in two distinct structural states is not new. For decades, scientists have theorized that liquid water is composed of two interconvertible local structures – one denser and more disordered, the other less dense and more ordered. The concept traces back to early physical chemists who noticed that no simple molecular model could account for water’s full array of anomalous behaviors under varying conditions of temperature and pressure.
This “two-state model” has been invoked to explain water’s many anomalous properties, including why it becomes easier to compress as it cools and why it reaches maximum density at 4°C rather than at its freezing point. Most liquids behave in a straightforward, predictable way: as temperature drops, they contract, become denser, and eventually freeze. Water does the opposite below 4°C, expanding as it cools further. Without the two-state model, this behavior has no clean mechanistic explanation.
But the model has remained controversial because direct molecular-level evidence for the two structures has been elusive. The structural states that theorists proposed are not static or cleanly separated – they interconvert on timescales faster than most experimental instruments can resolve. One reason water’s molecular structure has been a mystery is that researchers largely relied on measurements such as local density and molecular energy to identify different forms of water. While useful, these quantities could not clearly separate the two proposed structures, leaving the debate unresolved even after years of simulations and experiments.
The 2026 Breakthrough: AI and 74 Million Molecular Snapshots
The key advance in the 2026 Nature Physics study was methodological. The problem is not just experimental, according to the researchers. Even in simulations, traditional methods that measure local density and energy differences between molecules failed to cleanly separate the two structures. What was needed was a way to let the data reveal the hidden molecular fingerprint of each structure – without any human assumptions about what that fingerprint should look like.
A team led by Prof. Xiao Cheng Zeng of the City University of Hong Kong turned to AI. They trained an unsupervised neural network on 74 million simulated water configurations, letting it hunt for hidden patterns without human bias. The unsupervised approach was deliberate: rather than feeding the algorithm a predefined idea of what two structures should look like, the researchers allowed the model to find its own organizational logic within the data. As Zeng put it: “It is practically impossible for humans to intuitively guess or manually construct such complex, nonlinear parameters. We need AI’s help.”
The AI identified two distinct clusters – Structure A (dense and disordered) and Structure B (light and ordered) – across a wide range of temperatures and pressures. Critically, the model found that these two structures exist not as stable, isolated phases but as dynamically interchangeable states, constantly converting from one form to the other at the molecular scale. By employing unsupervised deep learning on massive molecular dynamics simulations, the team found that near the high-density and low-density phase boundary, interconvertible reactions proceed via a full-loop reaction pathway with three saddle points, whereas away from it the reactions proceed via a semi-loop pathway with a single saddle point. In plain terms: the path water molecules take when switching between the two structural forms changes depending on the thermodynamic conditions – a finding that itself has implications for understanding phase transitions more broadly.
These findings provide molecular-level evidence in support of the two-state water model and may offer physical insights into the origin of liquid-liquid phase transitions more generally.
The Two Structures: HDL and LDL
The two states identified by the AI correspond to forms that theorists had long predicted. One is a dense, disordered high-density liquid, or HDL. The other is a more open, ordered low-density liquid, or LDL. Both exist simultaneously in ordinary liquid water at room temperature, with individual molecules constantly switching between the two configurations.
Understanding what drives that switching requires a brief look at water’s molecular architecture. Each water molecule – one oxygen atom bonded to two hydrogen atoms – carries an unequal charge distribution. The oxygen-hydrogen bond in water is a polar covalent bond, giving each hydrogen atom a slight positive charge and the oxygen a slight negative charge. This polarity allows water molecules to form hydrogen bonds with their neighbors: the slightly positive hydrogen of one molecule is attracted to the slightly negative oxygen of another. Hydrogen bonds in water are weak interactions, approximately one-twentieth the strength of the oxygen-hydrogen covalent bond, and they rapidly form and break at room temperature.
Each water molecule can theoretically form up to four of these hydrogen bonds simultaneously, producing a local tetrahedral arrangement – four neighboring molecules arranged roughly like the points of a pyramid. In the low-density structure (LDL), this tetrahedral arrangement is relatively intact, producing an open, ordered, and less dense molecular configuration. In the high-density structure (HDL), that local geometry is more distorted and disordered, allowing water molecules to pack together more tightly.
The constant interconversion between these two states – HDL and LDL – at sub-picosecond timescales (one picosecond is one-trillionth of a second) is, according to the two-state model, what produces water’s anomalous thermodynamic profile. As temperature drops, the equilibrium shifts progressively toward the more ordered LDL configuration, which is why water becomes less dense as it approaches freezing, rather than more dense as most liquids do.
The Liquid-Liquid Phase Transition and the Second Critical Point
A central prediction of the two-state model that the 2026 findings support is the existence of a liquid-liquid phase transition (LLPT) in deeply supercooled water – water that has been cooled below 0°C without freezing. Central to the two-state model is a hypothesized phenomenon known as the liquid-liquid phase transition. The idea is that in the deeply supercooled regime, water splits into two macroscopically distinct liquid phases: a high-density liquid and a low-density liquid. The boundary between them is thought to terminate at a “second critical point.”
This second critical point is distinct from water’s well-known critical point (the temperature and pressure above which liquid and gas become indistinguishable). The proposed LLPT critical point sits in a deeply supercooled, high-pressure regime that is extraordinarily difficult to probe experimentally. The proposed transition between the two forms is believed to occur in deeply supercooled water, a state that is extremely difficult to study because water rapidly crystallizes into ice before scientists can observe it.
A separate 2026 study published in the Proceedings of the National Academy of Sciences advanced this picture further. Using microsecond-long molecular dynamics simulations with a highly accurate polarizable water model, researchers provided direct evidence of a well-defined moving interface between low-density and high-density water near the phase boundary – a finding consistent with a first-order transition between two liquid phases separated by a genuine free energy barrier.
The convergence of these independent lines of computational evidence in 2026 represents a qualitative shift in confidence among researchers. Multiple teams, using different simulation approaches and water models, are now reaching consistent conclusions about the existence and location of the liquid-liquid transition – even as experimental confirmation at those conditions remains a formidable unsolved problem.
Why Water’s Anomalies Matter Beyond the Lab
The distinctive characteristics of water, evident in its thermodynamic anomalies, have implications across disciplines from biology to geophysics. The most immediate consequence of water’s density behavior is one that every lake-dwelling species depends on: ice floats.
Ice is approximately 8 to 9 percent less dense than liquid water, with a density of around 917 kg/m³ compared to liquid water’s approximately 1,000 kg/m³. That density difference is a direct consequence of what happens to water’s molecular structure at freezing: the tetrahedral hydrogen-bond network locks into a fixed open hexagonal lattice that occupies more volume than the dynamic, partially disordered liquid. Because ice is less dense, it floats. Because it floats, it forms an insulating layer over lakes and rivers in winter, allowing liquid water – and the life it supports – to persist beneath. A world where ice sank would be a world where bodies of water froze solid from the bottom up, with consequences for aquatic ecosystems that are difficult to overstate.
The high cohesion between water molecules gives it a high freezing and melting point, such that we and our planet are bathed in liquid water. The large heat capacity, high thermal conductivity, and high water content in organisms contribute to thermal regulation and prevent local temperature fluctuations, allowing living things to more easily control body temperature. Each of these properties traces, at least in part, back to the same underlying structural dynamics that the 2026 Nature Physics study mapped for the first time at the molecular level.
Researchers are also exploring whether the two-state dynamics of liquid water play a role in the behavior of water inside living cells and around biological macromolecules. Water in the vicinity of proteins and DNA does not behave the same as bulk water – its hydrogen-bonding network is perturbed, and its structural equilibrium shifts. If the two-state model is correct, then the balance between HDL and LDL water near biological surfaces could have measurable consequences for how enzymes fold, how membranes function, and how molecular machines inside cells operate.
What Remains Open: Limits of the Current Evidence
The 2026 findings are grounded in computational simulations, not direct experimental measurement of real liquid water. That distinction matters. Molecular dynamics simulations use mathematical models – however accurate – to approximate the behavior of water molecules. The specific water model used in the Nature Physics study is widely regarded as one of the most accurate available, but no simulation model perfectly captures all of water’s properties across all conditions.
The deeply supercooled region where the liquid-liquid phase transition is predicted to occur is so hard to study experimentally because water crystallizes rapidly. Much of the evidence for the LLPT has therefore come from computational studies. The experimental challenge is not simply technical difficulty: it is that the relevant temperatures and pressures fall in a region where water cannot remain liquid long enough for measurements to be taken. Researchers are pursuing ultrafast X-ray scattering techniques and experiments on water confined in nanopores – environments where crystallization is suppressed – as potential routes to direct experimental access.
The next step, according to the research team, is deciphering the physical meaning of the AI-discovered structural traits and confirming them in real-world experiments. That confirmation step may take years, and it may require experimental methods that do not yet fully exist.
The broader field has also not reached unanimity. Some researchers favor alternative theoretical frameworks for explaining water’s anomalies that do not invoke a liquid-liquid transition. The two-state model now has its strongest computational support ever, but the debate is not closed.
You can explore more on how water’s molecular behavior affects your health in this Hearty Soul article on water and hydration science.
A Second Parallel Line of Research
The 2026 Nature Physics study was not the only significant computational advance on this question in recent years. A 2025 study, also published in Nature Physics, used a deep neural network to estimate the location of water’s liquid-liquid critical point at approximately 198 K (roughly minus 75°C) and 1,250 atmospheres of pressure. That estimate, derived from a different modeling approach, is broadly consistent with the ranges predicted by the 2026 work – lending further weight to the position that both a liquid-liquid transition and a second critical point exist, and that simulations are converging on their physical location.
Separately, a July 2026 study reported in Communications Chemistry used AI to compare 16 different structural descriptors for supercooled water, systematically identifying which parameters most efficiently distinguish high-density and low-density liquid states. The finding adds a methodological dimension to the broader research agenda: not only is water’s dual-structure behavior increasingly supported, but the scientific community is now developing more rigorous tools to characterize it.
Read More: Scientists make alarming discovery about health impact of drinking bottled water
Key Takeaways
The 2026 water forms discovery does not change what water is in any practical sense – the water coming out of your tap is the same substance it has always been. What it changes is the scientific understanding of why that substance behaves the way it does.
For the scientific community, the implications are substantial. If the two-state model is correct and the liquid-liquid phase transition is real, it means water has a previously unmapped phase diagram – a full region of its thermodynamic behavior that was theorized but not documented. Mapping that region experimentally, and understanding how the transition between HDL and LDL water operates at molecular resolution, will occupy researchers for years. The tools being developed to do that work – unsupervised machine learning applied to massive molecular dynamics datasets – are also tools that can be applied to understanding the behavior of other complex liquids, from molten metals to biological fluids.
For non-specialists, the clearest takeaway is that one of the most familiar substances in the world remains an active frontier of scientific inquiry. The 2026 findings represent not a conclusion, but a confirmation: water is genuinely stranger than it looks, the theoretical framework developed to explain that strangeness is increasingly well-supported, and the experimental work needed to fully validate it is just beginning. Understanding the hidden structural dynamics of water is not an abstract concern – it is foundational to understanding the conditions that make liquid water the universal medium of biology, the engine of Earth’s climate, and one of the most physically anomalous substances in the known universe.
AI Disclaimer: This article was created with the assistance of AI tools and reviewed by a human editor.