Ice Phase Classification Made Easy with Score-Based Denoising.

Area of Science:

• Materials Science
• Computational Chemistry
• Physics

Background:

• Accurate ice phase identification is crucial for understanding physicochemical phenomena.
• Classifying ice polymorphs in molecular dynamics simulations is challenging due to complex symmetries and thermal fluctuations.
• Existing methods often require expert knowledge, specific geometric data, or extensive training datasets.

Purpose of the Study:

• To develop an unsupervised phase classification framework for molecular dynamics simulations.
• To overcome limitations of traditional and existing machine learning approaches for ice phase identification.
• To provide a generalizable method for analyzing structural evolution in complex materials.

Main Methods:

• A novel unsupervised framework combining a score-based denoiser and a model-free classifier.
• Training the denoiser on perturbed synthetic data of ideal reference structures.
• Utilizing smooth overlap of atomic position (SOAP) descriptors for atomic fingerprinting, ensuring Euclidean symmetry and transferability.

Main Results:

• Achieved 100% accuracy in distinguishing ice phases using only seven ideal reference structures.
• Demonstrated the generalizability of the score-based denoiser model for phase identification.
• Eliminated the need for large datasets and manual labeling efforts.

Conclusions:

• The proposed framework offers an accurate and efficient method for ice phase identification.
• The approach is broadly applicable to various materials, aiding in the study of structural evolution.
• Provides new insights into the fundamental understanding of water and other complex molecular systems.