Enhanced Sampling of Crystal Nucleation with Graph Representation Learnt Variables.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Deriving low-dimensional variables from complex crystal structure data is challenging.
• Enhanced sampling methods require accurate thermodynamic information for reliable predictions.
• Traditional methods may struggle with complex phase transitions and polymorphism.

Purpose of the Study:

• To develop a graph neural network (GNN)-based autoencoder for extracting meaningful low-dimensional variables from crystal structure features.
• To utilize these variables in enhanced sampling techniques for observing state-to-state transitions and calculating thermodynamic weights.
• To validate the GNN approach by examining the nucleation of iron and glycine polymorphs.

Main Methods:

• Utilized a graph neural network (GNN) with autoencoder architecture.
• Employed simple convolution and pooling operations within the GNN.
• Integrated derived graph latent variables into well-tempered metadynamics for enhanced sampling.
• Applied the protocol to study nucleation in molten iron and glycine allotropes/polymorphs.

Main Results:

• The GNN-derived latent variables successfully captured essential features of crystal structures.
• Biased enhanced sampling using GNN variables consistently revealed state-to-state transitions.
• Accurate thermodynamic rankings were achieved, aligning with experimental data.
• The method demonstrated dependable sampling, crucial for materials discovery.

Conclusions:

• The proposed GNN-based approach offers a powerful tool for enhanced sampling in materials science.
• Graph latent variables provide reliable insights into material behavior and phase transitions.
• This protocol shows promise for broader applicability across different systems and sampling methods.