Machine learning coarse grained models for water.

Area of Science:

• Computational chemistry
• Materials science
• Thermodynamics

Background:

• Accurate molecular-level simulation of ice-water systems at mesoscopic scales is computationally challenging.
• Existing atomistic models are computationally expensive for large-scale simulations.

Purpose of the Study:

• To develop computationally efficient and accurate machine-learned coarse-grained (CG) models for ice-water systems.
• To describe mesoscopic behavior, structural properties, and thermodynamic anomalies of water and ice.

Main Methods:

• Introduced machine-learned coarse-grained (ML-CG) models: ML-BOP, ML-BOPdih, and ML-mW.
• Employed a multilevel evolutionary strategy for training CG models.
• Incorporated first-principles energetics, experimental data, and temperature-dependent properties from molecular dynamics.

Main Results:

• Achieved accurate description of water and ice structure and thermodynamic anomalies at mesoscopic scales.
• Models demonstrated a two orders of magnitude reduction in computational cost compared to atomistic models.
• ML-BOP models correctly predicted the experimental melting point of ice and the temperature of maximum density of liquid water.

Conclusions:

• The developed ML-CG models offer a significant advancement in simulating ice-water systems efficiently and accurately.
• The novel training strategy and ML workflow improve upon existing CG models.
• These models enable more accessible and detailed studies of mesoscopic phenomena in water and ice.