Distinguishing liquid crystalline nematic variants by machine learning.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Machine Learning

Background:

• Liquid crystals exhibit diverse phases, including ferroelectric and twist-bend nematic.
• Distinguishing these phases is crucial for understanding their properties and applications.
• Machine learning offers potential for automated phase identification.

Purpose of the Study:

• To evaluate sequential convolutional neural networks (CNNs) and parallel inception models for identifying nematic liquid crystal variants.
• To determine the optimal model complexity and data augmentation strategies for accurate phase classification.
• To assess the performance of different machine learning architectures in distinguishing ferroelectric and twist-bend nematic phases.

Main Methods:

• Trained CNNs (1-5 layers) and inception models (1-3 blocks) on liquid crystal data.
• Applied data augmentation techniques including flip, contrast, and brightness.
• Incorporated dropout-layer regularization to prevent overfitting.
• Systematically analyzed accuracy and error rates for each model configuration.

Main Results:

• Flip augmentation significantly improved accuracy, while dropout regularization generally decreased it.
• A 3-layer CNN or a single inception block model achieved accuracies of 0.96-0.98 ± 0.01 with flip augmentation.
• Sequential CNNs are sufficient for classifying sequences with up to four nematic phases.
• Higher accuracies approaching 100% are achievable with larger, class-balanced datasets.

Conclusions:

• Machine learning models, particularly 3-layer CNNs, can reliably distinguish nematic liquid crystal variants.
• Flip augmentation is a key technique for enhancing classification accuracy.
• For datasets with four or fewer phases, sequential CNNs offer a computationally efficient solution.
• Inception models may offer benefits for larger datasets, provided overfitting is managed.