A Machine Learning Model to Classify Dynamic Processes in Liquid Water.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Machine Learning Applications

Background:

• Understanding water dynamics is crucial for various scientific fields.
• Hydrogen bonds (H-bonds) significantly influence water's properties and behavior.
• Distinguishing dynamic processes like H-bond interchange and breakage is challenging.

Purpose of the Study:

• To investigate the dynamics of hydrogen bonds in bulk water using computational and machine learning methods.
• To classify and quantify H-bond interchange and breakage events.
• To explore the potential of deep learning in analyzing complex molecular dynamics.

Main Methods:

• Ab initio molecular dynamics simulations were employed to generate trajectory data.
• A novel directed Hydrogen (H-) bond population operator was defined.
• A recurrent neural network (RNN)-based deep learning model was designed and trained.
• The model was used to classify H-bond interchange and breakage events.

Main Results:

• The study successfully classified H-bond interchange and breakage processes in water.
• The ratio of interchange to breakage events was determined to be approximately 1:4.
• This ratio remained consistent across a temperature range of 280–360 K.
• The findings highlight the distinct dynamics of H-bond interchange and breakage.

Conclusions:

• Deep learning models can effectively distinguish complex dynamic processes involving hydrogen bonds.
• The 1:4 interchange-to-breakage ratio provides quantitative insight into water's H-bond dynamics.
• This approach has broad applicability for studying H-bond dynamics in other molecular systems.