IceCoder: Identification of Ice Phases in Molecular Simulation Using Variational Autoencoder.

Area of Science:

• Computational physics and chemistry
• Materials science
• Machine learning applications

Background:

• Classifying diverse ice phases in molecular simulations is challenging due to complex phase space and thermal fluctuations.
• Traditional order parameters are often insufficient for differentiating subtle phase variations.

Purpose of the Study:

• To develop a novel machine learning framework, IceCoder, for accurate and efficient identification and classification of ice phases.
• To overcome limitations of traditional methods in distinguishing ice polymorphs and liquid water.

Main Methods:

• Utilized a variational autoencoder (VAE) combined with smooth overlap of atomic position (SOAP) descriptors.
• Compressed high-dimensional SOAP vectors into a 2D latent space for visualization and classification.
• Trained the model on data from molecular dynamics simulations.

Main Results:

• IceCoder effectively classifies various crystalline ice phases and liquid water at the molecular level.
• The 2D latent space facilitates clear distinction between different ice phases.
• Demonstrated robust and generalizable performance in tracking ice phase transitions.

Conclusions:

• IceCoder offers a powerful and computationally efficient tool for analyzing ice phases in simulations.
• The framework can be generalized to study polymorphs in other molecular crystals.
• Provides new insights into microscopic mechanisms of nucleation, growth, and phase transitions.