💧 Water would be made of two liquids in one

Everyone knows that water has unusual behavior: it becomes less dense when freezing, allowing ice to float. But this singularity is only the visible part of a much larger phenomenon.

In reality, water exhibits dozens of surprising properties, such as an abnormally high heat capacity compared to similar liquids. For decades, scientists have believed that these anomalies might have a common origin: at the molecular scale, water would actually be two distinct liquids that constantly interconvert.


This hypothesis, called the "two-state" hypothesis, posits the existence of a dense form and a less dense form of water, which coexist and convert into each other. Proving its existence is extremely difficult, because the transformation is so fast and subtle that it escapes conventional measurements. But a team of researchers, using artificial intelligence, has just taken an important step by providing the first direct molecular evidence of this phenomenon, as they report in Nature Physics.

To achieve this, the team led by Xiao Cheng Zeng used an innovative approach: unsupervised deep learning. Instead of looking for pre-defined patterns, the AI was trained to analyze molecular dynamics involving hundreds of thousands of water molecules. By extracting the most relevant variables, it identified how a water molecule transitions from a dense structure to a loose structure, and vice versa. This method allowed mapping the process in record time, where traditional approaches would have required years.

The results reveal that the conversion between the two forms follows two distinct paths. Most often, it follows a trajectory with a single energy barrier to overcome. But near the boundary where the two states coexist – a threshold similar to that between ice and liquid water – molecules can take a longer path, which involves three successive barriers. This discovery explains why water seems to change behavior under extreme conditions, for example when it is supercooled.

These results do not just validate a thirty-year-old hypothesis. They could also help understand many properties of water, such as its abnormal density, viscosity, or heat capacity. By better understanding the molecular structure of water, researchers hope to eventually improve modeling of interactions with salts, proteins, and drugs in solution, which would have applications in pharmacy and biology.

But there is still work to be done. Xiao Cheng Zeng and his team are currently developing a more powerful machine learning model to confirm these results. The next step will be to validate them experimentally, which requires extremely sensitive measurement techniques. If this work succeeds, it could shed light on one of the greatest mysteries in physics: why is water so singular?