Researchers at the University of Osaka have applied artificial intelligence to assess methods for describing molecular structures in supercooled water. The study, published in Communications Chemistry, uses a neural network to compare structural descriptors and evaluate their effectiveness.
Water exhibits unusual properties, such as expanding when frozen, which are connected to changes in its microscopic structure under varying temperature and pressure. No standard system exists to classify these changes. Supercooled water remains liquid below its freezing point in clean containers without nucleation sites.
Anomalies intensify during supercooling and have been linked to shifts between high-density and low-density liquid states. These states arise from evolving networks of hydrogen bonds, with denser structures becoming more common as temperature rises.
The team trained a neural network on data from molecular dynamics simulations. The model tested 16 descriptors, including tetrahedral order and local density, to distinguish between the two liquid states. Results identified the most effective descriptors for capturing structural differences.
The approach may improve understanding of water’s thermodynamic behavior and support development of better characterization tools.
