A First Principles Investigation and Microscopic Classification of Water Glasses

The molecular origins of water's anomalous properties, particularly pronounced at low temperatures, have long puzzled researchers. The liquid-liquid transition (LLT) hypothesis, proposing distinct low-density (LDL) and high-density (HDL) liquid states with a transition terminating at a liquid-liquid critical point (LLCP), offers a compelling explanation with increasingly more experimental support. However, rapid crystallization near the LLCP has shifted focus to water's glassy states: low-density amorphous (LDA) and high-density amorphous (HDA) ice. This study employs the Deep Potential (DP) framework, leveraging machine learning for first-principles accuracy at unprecedented efficiency, to investigate water's behavior in supercooled and glassy regimes. Using two DP models, DP-SCAN (based on density functional theory) and DP-MBpol (using the Many-Body Polarizable water model), we demonstrate their ability to accurately extrapolate to glassy phases despite not being explicitly trained on them. Our results show a first-order-like LDA-HDA transition upon isothermal compression, with density changes and structural differences between the glasses aligning well with experimental data. We then demonstrate microscopic analysis via a neural network classification algorithm, which provides a detailed picture of the LDA/HDA transition. We compare thermodynamic and microscopic calculations of LDA/HDA spinodals in the pressure-temperature plane, and calculate distinct glass transition temperatures for LDA and HDA from isobaric quenches, suggesting these glasses serve as proxies for their corresponding liquids. This work advances our understanding of water's complex phase behavior, demonstrates the potential of data-driven approaches in exploring challenging states of matter, and offers new avenues for investigating complex molecular systems with unprecedented fidelity and efficiency.