Researchers at the University of Osaka have used artificial intelligence to compare models of water's microscopic structure in its supercooled state. The work identifies the most effective descriptors for distinguishing between two competing liquid forms of water.
Scientists have long observed that water expands when it freezes and displays other unusual behaviors. These traits are linked to shifts between a high-density liquid and a low-density liquid at the molecular level, especially when water is cooled below its normal freezing point without solidifying.
The team trained a neural network on data from molecular dynamics simulations. The model evaluated 16 structural descriptors to determine which best captured key differences between the two liquid states at varying temperatures.
Corresponding author Kang Kim noted that the approach draws on machine learning methods shown to be effective for classifying structural data. Senior author Nobuyuki Matubayasi said the network determined the most efficient descriptors through this process.
The findings were published in Communications Chemistry. The framework may improve understanding of how molecular changes relate to water's thermodynamic properties.
