Water is the most common liquid on the Earth's surface, and it differs significantly in that it expands when it freezes. Water's anomalies are related to how its microscopic structure changes depending on temperature and pressure. However, there is no systematic framework for describing these structural changes.

Researchers from Osaka University are using artificial intelligence (AI) to evaluate systems for defining the characteristics of water. The AI model is part of a unified system for comparing and assessing structural descriptors of supercooled water. This exciting discovery was reported in Communications Chemistry.

For water to freeze, the molecules must become ordered and form a crystal lattice, as in ice. The molecules attach to a base, the so-called nucleation center, to form a solid phase. Impurities in the water or scratches on the inner surface of the container can serve as nucleation centers.

Therefore, water in a smooth, clean container can be cooled below the freezing point, but it will not turn into ice, reports Nauchnaya Rossiya. This state is called supercooling of water.

The anomalous behavior of water becomes more pronounced during supercooling. These anomalies are explained by a transition between two competing states: high-density liquid (HDL) and low-density liquid (LDL). At the microscopic level, order in water arises from a network of intermolecular hydrogen bonds that changes continuously over time. As the temperature rises, the compact HDL structures begin to dominate over the open LDL structures.

Various structural descriptors, such as tetrahedral bond order and local density, have been introduced to characterize local order in water. Since these descriptors were proposed independently of one another, they differ qualitatively in parameters and scales and encode different structural information. This makes it difficult to compare the descriptors systematically in order to assess their relative importance.

“Previous studies have shown that using machine learning to classify and understand structural data is effective,” explains study author Kang Kim. “We specifically wanted to include a neural network model in this study to evaluate how accurately the descriptors convey key structural information.”

The network was fed structural data on supercooled water obtained through computer experiments known as molecular dynamics simulations. To enable the network to recognize patterns in the data, it used a trial-and-error method.

“The network used the acquired knowledge to compare how 16 descriptors distinguished low- and high-density lipoprotein structures at different temperatures,” says Nobuyuki Matubayasi, the study’s senior author. “This allowed us to identify the most effective descriptors.”

The results obtained will help better understand the relationship between structural fluctuations and the thermodynamic states of water, determine the origin of water’s anomalous properties, and develop improved structural descriptors.