AI model pinpoints descriptors distinguishing water’s two liquid states

Water’s unusual behavior is magnified when it is supercooled, yet scientists have long struggled to compare the many microscopic descriptors that aim to capture its structure. A team at the University of Osaka has tackled this problem by applying an artificial‑intelligence model trained on extensive computer simulations. The AI was tasked with evaluating 16 distinct structural descriptors that have been proposed to characterize liquid water at the molecular level.

The analysis revealed which descriptors most effectively differentiate between the two competing liquid states that water can adopt under supercooled conditions. By ranking the descriptors according to their predictive power, the researchers have established a clearer framework for future studies of water’s complex phase behavior. The findings streamline the selection of structural metrics for both experimental and theoretical investigations, advancing the broader effort to understand one of nature’s most enigmatic substances.