Toward automated classification of monolayer versus few-layer nanomaterials using texture analysis and neural networks

The rapid advancement in two-dimensional (2D) materials, particularly transition metal dichalcogenides (TMDs) like MoS2 and WS2, has spurred a need for efficient characterization techniques. These materials show promise in various applications due to their unique electronic, optical, and mechanical properties. However, controlling the number of layers during synthesis is crucial for achieving desired material properties. Current methods like photoluminescence (PL) and Raman spectroscopy are time-consuming and may not be suitable for large-scale analysis. This research aims to develop a fast, robust, and automated method for distinguishing monolayer from few-layer TMDs using image analysis techniques, specifically focusing on texture analysis and machine learning.

Existing automated layer identification methods for 2D materials primarily rely on pixel intensity analysis (grey levels or RGB values). These methods suffer from limitations like sensitivity to acquisition conditions and insufficient consideration of pixel interrelationships. Studies by Jessen et al., Masubuchia, Xiaoyang Lin et al., Yuhao Li et al., and Lei et al. have explored using RGB spectral fingerprints, optical feature extraction with machine learning, and machine learning-assisted optical microscopy for material identification. However, these methods often neglect the spatial relationships between pixels, which contain valuable information about the material's texture. This research proposes to overcome these limitations by employing texture analysis, which is a robust image analysis technique considering pixel interdependencies.

The study used chemical vapor deposition (CVD) to grow MoS2 and WS2 samples. Optical microscopy images were acquired, and Raman and PL spectroscopy were used to confirm monolayer and few-layer regions. Three TA methods were applied to the optical images: GLH (first-order), GLCOM (second-order), and RLM (higher-order). The Fisher coefficient was used for automatic feature selection, choosing the top 10 (or 9 for GLH) features for each TA method. Two classifiers, ANN and LDA, were then used to classify the image regions as either monolayer or few-layer. The ANN architecture consisted of an input layer (number of nodes corresponding to the selected features), a hidden layer with 1 and 2 neurons and an output layer with 2 nodes for classification. The performance of each TA method combined with each classifier was evaluated based on classification accuracy.

The RLM method combined with the ANN classifier yielded the highest classification accuracy: 89% for MoS2 and 95% for WS2. GLCOM with LDA achieved high accuracy only for MoS2 (89%). GLH consistently showed lower accuracy. The results suggest that higher-order TA methods (like RLM) better capture the textural differences between monolayer and few-layer structures. The study found that ANN performed better than LDA when using GLH and RLM features, while LDA performed better with GLCOM features. Tables 1 and 2 list the best features selected for MoS2 and WS2, respectively, based on the Fisher coefficient. Table 3 summarizes the classification accuracy for each combination of TA method and classifier. The analysis showed that RLM features, which reflect texture granularity (through 'run' properties), are better characterizing properties than homogeneity for automated classification of nanolayers.

The superior performance of RLM with ANN highlights the importance of considering higher-order statistical information in characterizing nanolayer thickness. The difference in accuracy between the TA methods can be attributed to the sensitivity of lower-order methods to variations in illumination and image acquisition. The higher-order methods are less sensitive to these variations. This approach offers a faster and potentially more reliable alternative to existing methods like Raman and PL spectroscopy. The results suggest that texture analysis coupled with ANN could be a powerful tool for automated quality control in the manufacturing of 2D TMD-based devices.

This study successfully demonstrated the automated classification of monolayer and few-layer MoS2 and WS2 using image texture analysis and neural networks. RLM with ANN achieved the highest classification accuracy. This approach provides a promising avenue for rapid and accurate characterization of 2D TMDs. Future work could focus on investigating the relationship between texture features and the underlying physical and chemical properties of the materials and explore the applicability of this method to other 2D materials and broader ranges of thicknesses.

The manual delineation of regions of analysis might introduce some subjective bias. Further research is needed to automate this process for improved objectivity. The current study focused on specific CVD-grown samples; the generalizability of the method to other growth techniques or material systems requires further investigation. The effect of defects and other imperfections on the accuracy of the classification warrants further examination.