Researchers at the University of Osaka have used an artificial intelligence model to systematically compare 16 structural descriptors for supercooled water, identifying the most efficient ones for distinguishing between high-density and low-density liquid states. The work, published in Communications Chemistry, provides a unified framework that could sharpen understanding of why water behaves unlike almost any other liquid.
Water expands when it freezes, and its anomalies grow more pronounced when it is cooled below the normal freezing point without solidifying - a condition known as supercooling. In a smooth, clean container, water can remain liquid well below 0°C because molecules lack nucleation sites to begin forming ice crystals.
Two liquid states in competition
At the microscopic level, supercooled water is thought to fluctuate between two competing arrangements: a high-density liquid (HDL) and a low-density liquid (LDL). The balance between these states shifts with temperature. As water warms, collapsed HDL structures dominate; at lower temperatures, open LDL structures become more prevalent. Precisely characterizing this local order has been challenging because existing descriptors - such as tetrahedral bond order and local density - were proposed independently, differ in dimension and scale, and encode different structural information.
Testing descriptors with a neural network
The Osaka team fed structural data from molecular dynamics simulations into a neural network and let it learn to classify patterns through trial and error. The network then evaluated how accurately each of 16 descriptors captured the differences between LDL and HDL structures across temperatures.
"Past studies have shown that using machine learning to classify and understand structural data is effective," said Kang Kim, corresponding author. "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 approach illustrates how AI for Science & Research can tackle classification problems where conventional comparisons fall short. Nobuyuki Matubayasi, senior author, said: "The network used what it had learned to compare how 16 descriptors differentiated between LDL and HDL structures at different temperatures. In this way, we determined the most efficient descriptors."
Sharper tools for water structure
The results give researchers a practical way to rank descriptors and select the ones most sensitive to the structural fluctuations that drive water's thermodynamic behavior. That could help pinpoint where water's anomalous properties originate and guide the design of improved descriptors for other complex liquids.
Why this matters for general, science and research
For scientists who simulate or analyze liquid-state systems, having a systematic way to evaluate structural descriptors removes guesswork and reduces the risk of missing subtle phase-relevant changes. The neural network method described here is not specific to water - it offers a template for benchmarking descriptors in any material where competing local structures influence macroscopic properties. Researchers who want to apply such techniques can deepen their skills through structured AI Research Courses that cover neural network analysis of physical science data.
Your membership also unlocks:
