Researchers Rank 16 Supercooled Water Descriptors, With LSI Scoring AUC Above 0.99
Updated
Updated · Nature.com · Jul 6
Researchers Rank 16 Supercooled Water Descriptors, With LSI Scoring AUC Above 0.99
1 articles · Updated · Nature.com · Jul 6
Summary
A neural-network benchmark of 16 structural descriptors found LSI and zeta best captured temperature-driven changes in supercooled water, with LSI reaching AUC of at least 0.991 under isochoric conditions and 0.963 under isobaric conditions.
Simulations covered 200-300 K for 1,000 TIP4P/2005 water molecules, using temperature classification as a proxy for how well each descriptor distinguishes HDL-like from LDL-like local structures.
Four descriptors—LSI, zeta, NTC HB and N HB—formed the top tier, showing that both coordination-shell separation and hydrogen-bond network connectivity are key to tracking structural crossover in supercooled water.
Isobaric runs at 1 bar generally outperformed fixed-density runs, especially at low temperatures near the roughly 230 K Widom-line crossover, indicating density fluctuations add useful structural information.
LIME explainable-AI analysis showed larger LSI and zeta values pushed predictions toward low-temperature, LDL-like states, giving the machine-learning rankings a physically consistent interpretation.
How will this AI insight help interpret the two liquid states recently confirmed in real-world experiments?
Water's dual nature is now clear to AI. How might this hidden structure dictate the fundamental processes of life?
AI has ranked our best tools for describing water. Can it now invent a superior one of its own?
Benchmarking 16 Structural Descriptors: LSI Achieves Near-Perfect AUC in Supercooled Water Phase Classification
Overview
A groundbreaking study published in July 2026 used machine learning to systematically evaluate 16 structural descriptors, aiming to accurately distinguish between the high-density liquid (HDL) and low-density liquid (LDL) states of supercooled water. This unified benchmarking approach provided clearer insights into the complex structural fluctuations behind water’s unusual properties. The research highlighted the Local Structure Index (LSI) for its near-perfect ability to discriminate between these states, marking a significant advance in understanding water’s behavior. Overall, this comprehensive evaluation sets a new standard for analyzing and interpreting the structure of supercooled water.