Water is the most abundant liquid on Earth’s surface and remains uniquely anomalous because it expands upon freezing.
These anomalies stem from changes in its microscopic structure driven by variations in temperature and pressure, but scientists have lacked a systematic way to characterize these structural transformations.
Researchers at Osaka University recently employed artificial intelligence to assess different characterization frameworks. Their AI model forms part of a unified system for comparing and estimating structural descriptors of supercooled water, as reported in the journal Communications Chemistry.
For water to solidify, its molecules must arrange into an ordered lattice, such as ice. This process requires molecules to attach to a nucleation site—like an impurity or a scratch in a container—that serves as a foundation for crystal growth.
Water held in a smooth, clean container can be cooled below its freezing point without turning into ice, a phenomenon known as supercooling.
As water is supercooled, its unusual properties become more evident. This behavior can be explained by a transition between two competing states: high-density liquid (HDL) and low-density liquid (LDL).
At the microscopic level, a network of constantly reshaping hydrogen bonds maintains water’s structure. As temperatures rise, HDL structures become more prevalent than the open LDL configurations.
Scientists have introduced various structural descriptors, such as tetrahedral bond order and local density, to characterize the local water structure. Because these descriptors were proposed independently, they vary in dimensions, scales, and the types of encoded structural information.
This diversity complicates systematic comparisons to evaluate their relative significance.
“Past studies have shown that using machine learning to classify and understand structural data is effective,” corresponding author Kang Kim stated.
READ MORE: Here’s why the South Pole froze over before the North Pole
“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’s inputs consisted of structural data for supercooled water derived from molecular dynamics simulations. It used a trial-and-error method to identify patterns.
Senior author Nobuyuki Matubayasi explained, “The network utilized its learned knowledge to compare how 16 descriptors distinguished between LDL and HDL structures across various temperatures.”
This process helped determine the most effective descriptors. The results could enhance scientific understanding of how structural fluctuations relate to water’s thermodynamic states.
These insights may also reveal the origins of water’s unusual properties and guide the development of better structural descriptors.