An interesting study recently published in the journal Nature Physics by Liwen Li and others titled “Evidence for the generic existence of two local structures in liquid water” offers compelling evidence supporting a long-debated theory about one of the most abundant molecules - water ($H_2O$).

We know that water behaves differently compared to other liquids. Water exhibits anomalous behavior as it has its maximum density at 4℃ and $1 \ atm$, meaning, it expands as the temperature decreases below 4℃. Studying this behavior, scientists came up with a theory that water is actually a mixture of two distinct local structures which are separated by a phase boundary. These are called:

  • Low Density Liquid (LDL) Water
  • High Density Liquid (HDL) Water

However, observing these phases has been difficult historically. It was believed that these phases could only be separated at the liquid-liquid critical point, which was estimated to be $198K$ and $1250 \ atm,$ in extreme supercooled conditions. 

But in this new study, the researchers used unsupervised deep learning techniques on molecular dynamics simulations an accurate and widely used water model. With this, they could map the exact molecular mechanics of how water transitions between LDL and HDL phases. 

They discovered that even though the transition is complex near the critical point, it simplifies as we move away from it. This study proves that these two phases are a fundamental property of liquid water across varying conditions, not just an anomaly at extreme conditions.

This study demonstrates how AI and deep learning can accelerate scientific research that was previously considered to be out of reach.

Link to the article: https://www.nature.com/articles/s41567-026-03301-8