Carbon nanotubes (CNTs), with their nanoscale diameter and high carrier mobility, are promising materials for next-generation energy-efficient electronic systems. Their applications range from low-power chemical and physical sensing to the fabrication of entirely CNT-based microprocessors. However, the commercialization of CNT-based electronics requires rapid characterization methods for quality control throughout large-scale fabrication. Raman spectroscopy offers a non-destructive, non-contact approach with micrometer spatial resolution, providing chemical and structural information. The G band shape and resonant electronic Raman scattering are crucial for distinguishing CNT types. However, the low efficiency of Raman scattering often leads to low signal-to-noise ratios (SNRs), making high-speed Raman imaging challenging. Traditional Raman spectroscopy requires high power and long exposure times, hindering high-throughput analysis. Machine learning (ML), particularly deep learning, offers a potential solution. Deep neural networks, especially Convolutional Neural Networks (CNNs), excel at pattern recognition and have shown success in analyzing spectroscopic data. This work focuses on developing a high-throughput approach to rapidly identify suspended CNTs by combining deep learning, high-speed Raman spectroscopy, and an optimized scan strategy. The suspended CNTs are grown on fork-like substrates optimized for high-throughput dry transfer, a process proven for clean and CMOS-compatible CNT integration in various applications. The workflow involves high-speed Raman line scans, followed by individual spectra classification using a trained deep learning model and a final optimized scan strategy to identify position, quantity, and metallicity of CNTs.

The introduction adequately cites relevant literature on CNT applications in electronics, the use of Raman spectroscopy for CNT characterization, and the application of machine learning in similar contexts. The authors highlight the challenges of low SNRs in high-speed Raman imaging and the potential advantages of employing deep learning for overcoming these limitations. Several papers are referenced regarding the use of Raman spectroscopy in characterizing CNTs, highlighting its advantages and limitations. The use of machine learning in various fields such as image and speech recognition is also discussed, laying the groundwork for the application of deep learning in this specific context of high-throughput CNT identification.

The authors created datasets of Raman spectra acquired under various conditions (different integration times and laser powers) to train and validate their deep learning model. The integration times ranged from 1 to 600 ms at 1 mW power, and laser powers varied from 0.1 to 3 mW for a fixed integration time of 50 ms. A large-scale labeled dataset (62,130 spectra) was constructed, categorizing spectra as M-CNTs (metallic), S-CNTs (semiconducting), or empty. A separate validation dataset (48,887 spectra) was also used. The signal-to-noise ratio (SNR) was quantified for each dataset and its relationship to integration time and laser power was investigated. A convolutional neural network (CNN) architecture was selected through a grid search. The optimal CNN consisted of 4 convolutional layers, followed by fully connected layers, a softmax layer (for classification probabilities), and a cross-entropy loss function. The model’s performance was evaluated based on accuracy across different integration times and laser powers. The effect of varying training and validation parameters was also investigated. To enhance the classification accuracy, a threshold was applied to the output of the softmax layer to reduce false positives and negatives. Finally, an optimized scanning strategy (rescanning areas of interest) improved the efficiency of CNT localization and identification compared to simple line scans.

The study achieved high classification accuracy in identifying suspended CNTs using high-speed Raman imaging and deep learning. Even with low SNRs (0.9), the accuracy exceeded 90%, reaching 98% for SNRs of 2.2. Applying a threshold to the softmax layer output further improved accuracy. The optimized CNN model showed robustness against variations in integration times and laser powers used during testing compared to training. Models trained on lower SNR data generally performed well on higher SNR data, while the reverse was not true. Applying a threshold to the softmax layer effectively reduced false positives while maintaining a high number of true positives. This thresholding significantly improved the reliability of the rapid Raman mapping, reducing the noise in classification by filtering out low-confidence predictions.

The findings address the research question by demonstrating a viable high-throughput approach for identifying CNTs. The combination of high-speed Raman imaging and deep learning significantly speeds up the characterization process compared to traditional methods, which rely on longer integration times and manual analysis. This is crucial for industrial adoption, where quality control during large-scale fabrication is essential. The robustness of the model against variations in acquisition parameters is a significant advantage, making it less susceptible to experimental drift. The ability to use the model effectively with low SNR data is also valuable, allowing for faster acquisition times. The optimized scanning strategy further enhances the method's efficiency and accuracy. The approach has the potential to be adapted for other nanomaterials, thus holding broad implications for nanomaterial characterization and quality control within manufacturing processes.

This research successfully demonstrated a high-throughput approach for rapid identification of suspended CNTs using a combination of high-speed Raman spectroscopy and deep learning. The method achieved high accuracy even under low SNR conditions, showcasing its robustness and efficiency. The use of a softmax threshold and optimized scanning strategy further enhanced performance. This approach has the potential for significant impact on nanomaterial quality control in industrial settings and is easily adaptable to other types of nanomaterials. Future research could explore the integration of this approach into fully automated production lines and examine its applicability to other types of nanomaterials and more complex nanomaterial systems.

The study primarily focused on suspended CNTs grown on specific fork-like substrates. The generalizability of the method to other CNT growth methods or substrates requires further investigation. While the model showed robustness to some variations in acquisition parameters, extreme deviations might still affect the accuracy. The current method relies on labeled data for training, which might be time-consuming for other materials. The study did not explicitly address the potential for variations in CNT properties beyond metallicity (e.g., diameter, chirality) and their impact on classification accuracy.