Water expands when it freezes, a quirk that makes it an outlier among liquids. Now scientists in Japan have used artificial intelligence to untangle why water behaves so strangely, especially when it is supercooled far below its freezing point.
Researchers at the University of Osaka trained an AI model on computer simulations of water molecules. The system evaluated 16 different ways of describing water's microscopic structure and identified which ones best distinguish between two competing liquid states. The findings appeared in Communications Chemistry.
Supercooled water hides two competing forms
Water can stay liquid even after it drops below 0 degrees Celsius if no impurities or rough surfaces are present to trigger ice formation. This state is called supercooled water, and its odd properties become even more extreme.
Scientists believe that supercooled water exists as a balance between two forms: high density liquid (HDL) and low density liquid (LDL). At the molecular level, water molecules constantly form and break hydrogen bonds. As temperature rises, the more compact HDL structures become dominant over the open LDL arrangements.
AI settles a long debate over measurement
For years, researchers proposed many ways to describe the local arrangement of water molecules, including tetrahedral bond order and local density. These structural descriptors were developed independently, using different scales and dimensions, making direct comparison nearly impossible.
Past studies showed that machine learning can classify and understand structural data effectively. The Osaka team specifically incorporated a neural network to evaluate which descriptors capture the most important features of water's structure. The AI system provided a unified framework for comparing these competing models.
Local people in Osaka and across Japan have long taken pride in the country's contributions to fundamental science. Water is essential to life and industry, and understanding its behavior at the molecular level could influence fields from climate modeling to materials science.
A clearer view of nature's most mysterious liquid
The AI's ability to rank structural descriptors gives scientists a reliable tool for studying water's dual liquid states. This does not solve every mystery about water, but it provides a consistent way to compare research methods that were previously incompatible. The work moves the field closer to a unified understanding of how water's microscopic structure drives its macroscopic oddities.
