Five-Site Water Models for Ice and Liquid Water Generated by a Series-Parallel Machine Learning Strategy.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Accurate molecular simulation of ice and liquid water is challenging due to limited model parameters.
• Understanding icing requires precise descriptions of water's molecular behavior.

Purpose of the Study:

• To develop accurate and versatile molecular models for water and ice using machine learning.
• To establish conventional TIP5P-BG and temperature-dependent TIP5P-BGT models.

Main Methods:

• A series-parallel machine learning approach was employed.
• This included classification back-propagation neural networks (BPNNs), parallel regression BPNNs, and a genetic algorithm.
• The models were trained to balance key physical properties of water.

Main Results:

• The developed TIP5P-BG and TIP5P-BGT models showed excellent agreement with experimental data.
• Mean absolute percentage errors were 2.65% and 2.40% for critical physical properties.
• Simulations of ice nucleus size and growth rate aligned well with experimental observations.

Conclusions:

• The study provides powerful molecular models for simulating phase transitions and icing in nanoconfinement.
• This work presents a novel strategy for constructing complex molecular models in extreme conditions.
• The findings advance the understanding of molecular-level icing processes.