Water in your glass would actually be two different liquids that continually interchange at the molecular scale. Researchers at the University of Hong Kong say they have found the first concrete molecular evidence of this transformation, thanks to an AI trained on tens of millions of data points. The results are published in Nature Physics.
What you will learn
- Why water behaves so differently from other liquids — and how the two-state hypothesis could explain it all
- How an unsupervised AI managed in about a year and a half what would have taken human researchers ten years
- Which two transition pathways water molecules take depending on the conditions
Water, a liquid with perplexing behaviors
Water is the most studied substance on the planet—and one of the most enigmatic. Unlike most liquids that become denser as they cool, water reaches its maximum density at 4°C, then expands. That is why ice floats. Water also resists temperature variations better than similar liquids, and its viscosity decreases under certain pressures. Dozens of anomalies have been documented without a satisfactory unifying explanation.
The two-state hypothesis offers such an explanation: at the molecular scale, water would consist of two structures — a high-density one and a low-density one — continuously interchanging. Fascinating on paper, this hypothesis has lacked direct molecular evidence for decades.
An AI trained to explore without direction
Xiao Cheng Zeng, a physical chemist at City University of Hong Kong, has studied water since the 1990s. Convinced by the two-state hypothesis but daunted by the complexity of proving it directly, he entrusted the problem about two and a half years ago to Liwen Li, a postdoctoral researcher in his laboratory.
Li proposed a radically different approach from conventional methods: unsupervised deep learning. An AI trained to detect structures in data without being told what to look for—forced to explore and create its own categories.
The team simulated the motions of hundreds of thousands of water molecules using the GROMACS software, generating tens of millions of data points. Without the AI, Zeng estimates the analysis would have taken about ten years. Li completed it in roughly eighteen months.
Two transition paths depending on the conditions
The AI produced “reaction coordinates” — a small set of variables that describe, with precision, how a water molecule shifts from a dense structure to a looser one, and vice versa. By tracing these transitions, the researchers found that the path taken varies with the conditions.
In most cases, the transition follows a so-called “semi-loop” route with a single energy barrier to cross. But near the boundary between the two states — analogous to the ice–water frontier at 0°C — the molecules can take a “full loop” path passing through three distinct barriers.
Zeng compares this duality to a mountain cut in two: a gentle slope on one side for most hikers, and near the shared summit, the possibility of making the full round.
Biological and pharmaceutical implications
Experimentally confirming this structure in real water will remain challenging and will probably require new, highly sensitive spectroscopic techniques. But if it is confirmed, the implications go beyond fundamental physics.
Nearly all biological and pharmaceutical processes occur in water. Understanding its true molecular structure could shed light on how salts, proteins, and drug molecules interact in solution — interactions essential to the functioning of injectable medicines and cellular processes.
👉 A small gesture for Space Group? Follow us on Google, add us to your bookmarks and find us on Google News so you don’t miss any discoveries.
