Beyond Local Structures in Critical Supercooled Water through Unsupervised Learning.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Statistical Mechanics

Background:

• The existence of a second critical point in water is a long-standing research question.
• Current models often rely on human-designed parameters to explain high-density (HD) and low-density (LD) water structures.

Purpose of the Study:

• To investigate the molecular origins of water's liquid-liquid critical point (LLCP) using unsupervised learning.
• To determine if distinct thermodynamic structures exist near the LLCP without human intervention.

Main Methods:

• Utilized atomistic simulations of water near the LLCP.
• Employed unsupervised machine learning with local and nonlocal descriptors to analyze structural environments.
• Compared results from local descriptors versus nonlocal descriptors probing nanometer-scale heterogeneities.

Main Results:

• Local descriptors did not reveal evidence for two distinct thermodynamic structures (HD and LD).
• Nonlocal descriptors, capturing nanometer-scale features, identified emerging LD and HD domains.
• These domains help rationalize the microscopic origins of density fluctuations near criticality.

Conclusions:

• The study challenges the necessity of predefined local structures to explain water's critical behavior.
• Nanometer-scale heterogeneities, revealed by nonlocal descriptors, are crucial for understanding density fluctuations near the LLCP.
• Unsupervised learning offers a powerful, data-driven approach to uncover complex phenomena in water simulations.