Computational investigation of water glasses using machine-learning potentials.

Area of Science:

• Physical Chemistry
• Computational Materials Science

Background:

• Water's anomalous properties are a long-standing scientific puzzle.
• The liquid-liquid phase transition hypothesis offers a framework for understanding these anomalies.
• Crystallization challenges have shifted focus to amorphous ice states.

Purpose of the Study:

• Investigate water's glassy phenomenology using quantum mechanical calculations.
• Assess the capability of machine-learning potentials in modeling amorphous ices.
• Explore the relationship between amorphous ice behavior and the liquid-liquid transition hypothesis.

Main Methods:

• Utilized two Deep Potential machine-learning models trained on DFT and MP potentials.
• Performed quantum mechanical calculations on water's glassy states.
• Simulated isobaric quenching and isothermal compressions of liquid water.

Main Results:

• Machine-learning models accurately captured amorphous ice structure and transformations, despite not being explicitly trained on them.
• Observed a continuum of amorphous ices and increased density fluctuations near the liquid-liquid critical pressure.
• Identified two distinct glass transition temperature branches for low-density and high-density amorphous ice, aligning with experimental data and the liquid-liquid transition hypothesis.

Conclusions:

• Machine-learning potentials trained on equilibrium phases can effectively model non-equilibrium glassy behavior.
• These models provide a powerful tool for studying long-timescale, out-of-equilibrium processes in water with quantum mechanical accuracy.
• The study supports the liquid-liquid phase transition hypothesis as a framework for water's anomalies.