Water is so familiar that its oddities are easy to overlook. Unlike almost every other liquid, it expands as it freezes, and it reaches its greatest density near 4 C rather than at its freezing point. For decades, scientists have traced these quirks to the way water's molecular structure rearranges as temperature and pressure change, yet they have lacked a shared, consistent way to measure and compare those rearrangements.
A team at the University of Osaka has now built one, with help from artificial intelligence. Writing in Communications Chemistry, Kohei Yoshikawa, Kokoro Shikata, Kang Kim and Nobuyuki Matubayasi describe a neural network that serves as a common benchmark for the many competing descriptions of water's local order.
The puzzle is sharpest in supercooled water, liquid chilled below its normal freezing point without turning to ice. Freezing needs a starting point, a nucleation site such as an impurity or a scratch inside a container; remove those, and water can stay liquid well below 0 C. In this state its anomalies grow more pronounced. Researchers explain them through a tug-of-war between two forms of the same liquid: a compact high-density liquid (HDL) and a more open, ordered low-density liquid (LDL), with hydrogen bonds constantly forming and breaking between molecules. As the temperature rises, the compact HDL arrangements increasingly win out.
To capture this, scientists have proposed many "structural descriptors", measures such as tetrahedral bond order and local density. Because each was devised on its own, they use different scales and dimensions and encode different information, which makes a head-to-head comparison difficult.
The Osaka group turned that comparison into a machine-learning task. They fed a neural network structural data from molecular dynamics simulations and trained it, through trial and error, to sort configurations by temperature. "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," said Kim, the corresponding author. The network then judged how well each of 16 descriptors separated LDL from HDL across temperatures, pinpointing the most efficient ones. A layer of explainable AI revealed which structural features drove its decisions.
Why it matters
The result is a data-driven yardstick for benchmarking descriptors, a step toward linking water's microscopic fluctuations to its measurable behavior. "The network used what it had learned to compare how 16 descriptors differentiated between LDL and HDL structures at different temperatures," said senior author Matubayasi. "In this way, we determined the most efficient descriptors." The framework, the team says, could sharpen the tools used to study water and help explain where its enduring anomalies come from.