Water expands when it freezes, a quirk that sets it apart from nearly every other liquid on Earth. Scientists have long suspected this and other odd behaviors trace back to shifts in water's microscopic structure, but they have lacked a consistent way to measure and compare those shifts. Now, a team at the University of Osaka has turned to artificial intelligence to close that gap, publishing its findings in Communications Chemistry.
The strangeness becomes most pronounced in supercooled water — liquid water that stays fluid below its normal freezing point because it lacks a nucleation site, a surface or impurity where ice crystals can begin forming. Without such a starting point, even tiny scratches inside a container or microscopic impurities, water can remain liquid at temperatures where it would normally freeze.
Researchers believe this behavior stems from a tug-of-war between two forms of liquid water at the molecular level: a high density liquid (HDL) and a low density liquid (LDL). Water molecules constantly form and break hydrogen bond networks, and as temperature rises, the more compact HDL arrangement increasingly dominates over the more open LDL structure.
Comparing Descriptors on a Level Playing Field
Over the years, scientists have proposed numerous ways to characterize the local arrangement of water molecules, including measures such as tetrahedral bond order and local density. But because these descriptors were developed independently of one another, they rely on different scales and types of information, making it hard to judge which ones actually capture the most meaningful structural detail.
To address this, the Osaka team trained a neural network on structural data generated from molecular dynamics simulations of supercooled water. Through repeated trial and error, the system learned to recognize meaningful patterns within the molecular structures.
"Past studies have shown that using machine learning to classify and understand structural data is effective," said corresponding author Kang Kim. "We specifically wanted to incorporate a neural network model into this study to evaluate how accurate the descriptors were at capturing key structural information, in a way that is like human cognition."
The network then used what it had learned to test how 16 different descriptors distinguished between LDL and HDL structures across a range of temperatures, according to senior author Nobuyuki Matubayasi. That comparison allowed the researchers to identify which descriptors most efficiently captured the essential structural differences.
The team says the resulting framework could deepen scientists' understanding of how microscopic structural changes connect to water's broader thermodynamic behavior. It may also help explain the roots of water's unusual properties and guide the development of better tools for probing its complex molecular structure in the future.