Water is caught in a structural dance between two competing forms
Water, the most familiar substance on Earth, has long concealed a structural secret within its own molecules — one that science has struggled to read clearly. Researchers at the University of Osaka have now enlisted artificial intelligence to evaluate competing methods of measuring water's microscopic behavior, particularly in its supercooled state where its strangeness is most pronounced. By training a neural network to rank the reliability of 16 different structural descriptors, the team has given science a more trustworthy lens through which to study a liquid that, despite its ubiquity, continues to defy easy understanding.
- Water's paradoxical tendency to expand when frozen has long pointed to something unusual happening at the molecular level — yet no reliable method existed to measure or compare those changes.
- Sixteen competing structural descriptors had accumulated over decades of research, each built on different scales and assumptions, making direct comparison nearly impossible and leaving the field without a common language.
- Scientists at Osaka trained a neural network on simulation data from supercooled water, letting the AI learn which descriptors most clearly distinguished between water's two competing molecular arrangements — high-density and low-density structures shaped by hydrogen bonds.
- The AI's rankings, published in Communications Chemistry in July 2026, do not merely tidy up a methodological dispute — they open a systematic path toward explaining water's thermodynamic oddities and studying other anomalous liquids.
Water covers most of our planet and anchors all known life, yet it remains one of science's most persistent puzzles. Its most famous quirk — expanding when it freezes — has long suggested that something unusual is happening at the molecular scale. The leading theory points to supercooled water, a state in which water remains liquid below its normal freezing point, as the place where this strangeness is most visible.
In that supercooled state, molecules are thought to oscillate between two competing arrangements: a tightly packed high-density structure and a more open low-density one, each shaped by the constant breaking and reforming of hydrogen bonds. As temperatures fall, the balance between these two forms shifts — and with it, water's behavior changes in ways that ripple up to the properties we observe every day.
The obstacle was not a lack of ideas but a lack of method. Over the years, researchers developed roughly 16 different ways to describe water's local molecular structure, each built independently with different scales and assumptions. Comparing them directly was nearly impossible, and no one could say with confidence which descriptor most accurately captured what was actually happening.
Kang Kim and colleagues at the University of Osaka turned to machine learning to resolve the impasse. They trained a neural network on simulation data from supercooled water, allowing it to learn which of the 16 descriptors most reliably distinguished between the two structural forms across a range of temperatures. The AI ranked them by effectiveness, identifying which measurements offered the clearest view into water's hidden architecture.
Published in Communications Chemistry in July 2026, the findings carry implications beyond methodological housekeeping. A more reliable framework for reading water's molecular grammar could help explain why it expands when frozen, why it boils at an unusually high temperature, and why it dissolves so many substances so well. For a liquid this central to life and this resistant to full understanding, a systematic way to study its inner workings feels long overdue.
Water is everywhere on Earth, and yet it remains one of science's most stubborn mysteries. It does something almost no other liquid does: it expands when it freezes. For decades, researchers have suspected that this bizarre behavior—along with water's other unusual properties—stems from changes in how its molecules arrange themselves at the microscopic level. But they've never had a reliable way to measure those changes or compare different methods of measurement. That gap has finally begun to close, thanks to work by scientists at the University of Osaka who enlisted artificial intelligence to decode water's hidden architecture.
The puzzle centers on what happens when water gets very cold. Normally, liquid water becomes ice when its molecules lock into an organized crystal lattice, a process that requires a starting point—a nucleation site. Tiny impurities or scratches inside a container can serve this purpose. But if those starting points don't exist, something strange occurs: water can stay liquid even after cooling below its normal freezing point. This state, called supercooled water, is where water's most exotic behavior emerges.
In supercooled water, scientists believe the molecules are caught in a kind of structural dance. The leading theory suggests the liquid oscillates between two competing arrangements: a high-density structure where molecules pack tightly together, and a low-density structure where they spread out more loosely. These configurations are created by the constant breaking and reforming of hydrogen bonds—the weak attractions that hold water molecules to one another. As temperature drops, the compact high-density form becomes more prevalent, while the open low-density form recedes.
The problem researchers faced was methodological. Over the years, scientists developed roughly 16 different ways to describe water's local molecular structure—measurements like tetrahedral bond order and local density. Each descriptor was created independently, using different scales and types of information. This meant comparing them directly was nearly impossible. Which one actually captured water's behavior most accurately? Which ones were redundant? There was no systematic way to know.
Kang Kim and his colleagues at Osaka decided to use machine learning to settle the question. They trained a neural network on structural data generated from computer simulations of supercooled water. The network learned, through repeated cycles of trial and error, to recognize meaningful patterns in the data. Then they asked it to evaluate all 16 descriptors—to measure how well each one distinguished between the high-density and low-density structures across different temperatures. The AI essentially ranked them by effectiveness, identifying which descriptors provided the clearest window into water's microscopic world.
The results, published in Communications Chemistry in July 2026, represent more than just a technical refinement. By establishing which structural measurements are most reliable, the framework opens a path toward deeper understanding of why water behaves the way it does. The researchers believe their approach could help explain the origins of water's thermodynamic oddities—why it expands when frozen, why it has such a high boiling point, why it's such an effective solvent. It may also lead to better tools for studying water itself, and perhaps other liquids that exhibit similarly anomalous behavior. For a substance that covers most of our planet and remains central to all known life, finally having a systematic way to read its molecular grammar feels overdue.
Citações Notáveis
We specifically wanted to incorporate a neural network model to evaluate how accurate the descriptors were at capturing key structural information, in a way that is like human cognition.— Kang Kim, corresponding author
The network compared how 16 descriptors differentiated between low-density and high-density structures at different temperatures, allowing researchers to determine the most efficient descriptors.— Nobuyuki Matubayasi, senior author
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that we can now rank these 16 different ways of measuring water's structure?
Because before, scientists were essentially using different languages to describe the same thing, and they couldn't tell which language was most accurate. Now we know which measurements actually tell us what's happening.
And what is happening, in the supercooled state?
Water is switching back and forth between two different arrangements—one dense, one loose. The AI helped us see that some descriptors capture this switching much better than others.
Does this explain why water expands when it freezes?
It's a piece of the puzzle. When water freezes, it's locking into the loose, open structure. That structure takes up more space. Understanding which measurements best track the shift toward that structure gets us closer to the full answer.
Could this method work for other liquids?
That's the real possibility. If we can use AI to evaluate structural descriptors for other anomalous liquids, we might unlock patterns we've been missing for years.
How did the neural network actually learn to rank these descriptors?
It was trained on simulations of supercooled water, then asked to distinguish between the two competing structures at different temperatures. Whichever descriptor did that job most reliably got ranked highest. It's like asking a student to grade different textbooks by how well they explain a concept.
And the researchers think this could improve our tools for studying water?
Yes. Once you know which measurements matter most, you can design better experiments and instruments focused on those specific measurements. You're not wasting effort on descriptors that don't actually tell you much.