Binary salt structure classification with convolutional neural networks: Application to crystal nucleation and...

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Accurate crystal structure classification is crucial for understanding material properties.
• Traditional methods can be computationally intensive and limited in scope.

Purpose of the Study:

• To develop and validate convolutional neural networks (CNNs) for classifying crystal structures of simple binary salts, specifically alkali halides.
• To assess the transferability and computational efficiency of CNN-based classification.

Main Methods:

• Utilized local bond orientational order parameters as input features for CNNs.
• Trained and validated CNNs on millions of data points from molecular dynamics (MD) simulations across nine bulk phases of alkali halides.
• Employed one-dimensional time convolution to filter structural fluctuations.

Main Results:

• Achieved high classification accuracy, up to 99.99%, on a large, balanced validation dataset.
• Demonstrated computational efficiency, with analysis being significantly less expensive than MD simulations.
• Successfully applied CNNs to track nucleation and crystal growth in alkali halide systems.

Conclusions:

• CNNs provide a highly accurate and computationally efficient method for crystal structure classification of alkali halides.
• The density-invariant input features enhance the transferability of the CNN classifiers.
• This approach offers a powerful tool for automated analysis in materials simulations, including melting point calculations.