Physically informed machine-learning algorithms for the identification of two-dimensional atomic crystals | JoVE Visualize

Area of Science:

Materials Science
Condensed Matter Physics
Computational Science

Background:

Graphene and other two-dimensional (2D) atomic crystals possess unique properties driving extensive research.
Mechanical exfoliation and optical microscopy are standard for fabricating and identifying 2D materials, but manual inspection is slow and impractical for scaling.
Deep learning (DL) methods automate 2D material identification but demand large datasets and use complex, uninterpretable algorithms.

Purpose of the Study:

To investigate the efficacy of transparent, tree-based machine learning algorithms for automating the identification of exfoliated 2D atomic crystals.
To compare the performance of tree-based models against a deep learning Convolutional Neural Network (CNN) in terms of accuracy and interpretability.
To analyze the decision-making processes of both tree-based models and CNNs to understand their reliance on physical image attributes.

Main Methods:

Exfoliated 2D materials (e.g., MoSe2) were analyzed using optical microscopy.
Tree-based machine learning algorithms (decision trees, gradient boosted decision trees, random forests) were trained using features mimicking color contrast.
A Convolutional Neural Network (CNN), specifically ResNet, was used as a benchmark for comparison.
Post-hoc analysis was performed to identify image regions utilized by the CNN for classification.

Main Results:

Tree-based algorithms successfully identified 2D atomic crystals by utilizing physical image features, specifically color contrast.
These models achieved high accuracy without the extreme overfitting or large dataset requirements typical of deep learning.
The study found that CNNs often rely on physically insignificant image attributes for classification, unlike the interpretable tree-based models.

Conclusions:

Tree-based machine learning offers a transparent, efficient, and data-frugal alternative to deep learning for automating 2D atomic crystal identification.
The interpretability of tree-based models allows for a better understanding of the physical basis of classification, crucial for scientific discovery.
Automating 2D material identification with interpretable AI can accelerate research and development in materials science.