Deep learning reveals a structural duality hidden within liquid water

Liquid water has long defied the standard thermodynamic expectations for simple fluids, growing more compressible as it cools and reaching its maximum density well before it freezes. For decades, a theoretical two-state model suggested these anomalies occur because water is not a uniform substance but a shifting mixture of two distinct microscopic structures. This model remained largely speculative because the rapid crystallization of water at low temperatures makes direct observation nearly impossible.

Using an unsupervised deep learning approach, researchers at the City University of Hong Kong analyzed 74 million molecular configurations from high-precision simulations. Unlike traditional methods that rely on human assumptions about density and energy, the team utilized an autoencoder to search for hidden mathematical fingerprints in the molecular data. By rotating the analytical "viewing angle" of these configurations, the model successfully isolated two distinct clusters of behavior.

The findings, published in Nature Physics, confirm the existence of a high-density liquid phase and a low-density liquid phase, with the latter showing a more ordered molecular arrangement. The study found that water molecules constantly interconvert between these two states. This structural duality provides a mechanical explanation for water’s unique properties, suggesting that the macroscopic behavior of the liquid is the result of a permanent internal tension between two different ways of being water.