Machine learning evaluation of structural descriptors for supercooled water

The anomalous behavior of liquid water is widely associated with a liquid-liquid phase transition between high- and low-density states in the supercooled regime. At the microscopic level, tetrahedral hydrogen-bond networks govern these properties, motivating structural descriptors that characterize local molecular environments. These structural descriptors quantify features such as tetrahedral order, local density, and the separation between the first and second coordination shells; however, they have largely been proposed independently, with limited systematic comparison. Here we evaluate 16 previously proposed descriptors using a neural-network-based temperature classification framework, enabling an objective assessment of their ability to distinguish temperature-dependent structural changes in supercooled water. We further apply an explainable artificial intelligence method that identifies the structural features responsible for the model predictions. This approach reveals how different descriptors encode local structural information and establishes a data-driven framework for benchmarking structural descriptors in liquid water. Structural descriptors for supercooled water have been largely proposed independently, differing in their definitions, dimensions and scales, but a systematic comparison is lacking. Here, the authors use a neural-network-based temperature classification framework to evaluate 16 previously proposed descriptors for their ability to distinguish temperature-dependent structural changes in supercooled water, and apply an explainable artificial intelligence method that identifies the structural features responsible for the model predictions.