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

TL;DR

This paper explores transparent, physically informed machine-learning algorithms like decision trees for identifying 2D atomic crystals in optical images, offering an interpretable alternative to deep learning methods.

Contribution

It introduces and compares tree-based machine learning methods that use physical image features for 2D crystal identification, emphasizing transparency and efficiency.

Findings

Tree-based algorithms successfully classify 2D crystal images.

These methods rely on physical image features for decision-making.

They avoid overfitting and require smaller datasets.

Abstract

First isolated in 2004, graphene monolayers display unique properties and promising technological potential in next generation electronics, optoelectronics, and energy storage. The simple yet effective methodology, mechanical exfoliation followed by optical microscopy inspection, used for fabricating graphene has been exploited to discover many more two-dimensional (2D) atomic crystals which show distinct physical properties from their bulk counterpart, opening the new era of materials research. However, manual inspection of optical images to identify 2D flakes has the clear drawback of low-throughput and hence is impractical for any scale-up applications of 2D samples, albert their fascinating physical properties. Recent integration of high-performance machine-learning, usually deep learning, techniques with optical microscopy has accelerated flake identification. Despite the advancement brought by deep learning algorithms, their high computational complexities, large dataset requirements, and more importantly, opaque decision-making processes limit their accessibilities. As an alternative, we investigate more transparent tree-based machine-learning algorithms with features that mimic color contrast for the automated identification of exfoliated 2D atomic crystals (e.g., MoSe2) under different optical settings. We compare the success and physical nature of the decisions of these tree-based algorithms to ResNet, a Convolutional Neural Network (CNN). We show that decision trees, gradient boosted decision trees, and random forests can successfully classify optical images of thin materials with transparent decisions that rely on physical image features and do not suffer from extreme overfitting and large dataset requirements.