High-speed identification of suspended carbon nanotubes using Raman spectroscopy and deep learning

Carbon nanotubes (CNTs) are promising candidates for next-generation, energy-efficient electronics due to their nanoscale diameter and high carrier mobility. They enable ultra-low-power sensing and have been used to realize a modern microprocessor based entirely on CNT field-effect transistors (CNFETs). For commercial adoption of CNT-based electronics, rapid characterization methods are needed for quality control during large-scale fabrication. Raman spectroscopy is a non-destructive, non-contact technique providing chemical and structural information with micrometer spatial resolution; for CNTs, features such as the G-band shape and resonant electronic Raman scattering distinguish metallic from semiconducting tubes. However, low Raman scattering efficiency necessitates higher power and long exposures, making high-speed imaging challenging, especially as subtle signals are easily masked by noise. Machine learning, particularly deep convolutional neural networks (CNNs), excels at pattern recognition and has shown promise for spectroscopic data analysis. This work introduces a high-throughput workflow combining high-speed Raman spectroscopy, deep learning-based spectral classification, and an optimized scan strategy to rapidly identify suspended CNTs grown across SiO2/Si fork-like substrates, enabling localization and classification into metallic (M-CNT), semiconducting (S-CNT), or empty regions with minimal acquisition time.

- Samples and scanning strategy: Individual CNTs are grown and suspended across SiO2/Si fork-like substrates optimized for high-throughput dry transfer. To maximize throughput, a line-scan strategy is implemented at each fork rather than full-area mapping. The line scan provides sufficient spectral information to localize CNT positions and classify them as M-CNT, S-CNT, or Empty. An optimized rescan strategy targets areas of interest to refine the position and amount of CNTs across each fork with higher efficiency than uniform line scans. - Data acquisition and SNR characterization: Raman spectra are acquired over a wide range of integration times (1–600 ms at 1 mW) and laser powers (0.1–3 mW at 50 ms) to study the speed–accuracy trade-off and determine optimal settings. Signal-to-noise ratio (SNR) is quantified and shown to scale linearly with both integration time and laser power; SNR < 1 occurs for integration times below 10 ms at 1 mW and at 0.1 mW with 50 ms. Example datasets include large sets of spectra from semiconducting and metallic CNTs acquired at multiple integration times and powers. - Dataset construction: Training dataset contains 62,130 labeled spectra across 19 Raman settings (integration time and power combinations), acquired on 21 trenches with M-CNTs, 20 trenches with S-CNTs, and 20 empty trenches. Validation dataset contains 48,887 labeled spectra acquired on 10 trenches with S-CNTs, 11 trenches with M-CNTs, and 15 empty trenches. The spectra exhibit significant spectrum-to-spectrum variability, especially at low integration times due to reduced SNR. - Neural network architecture and training: A CNN architecture is selected via grid search over candidate models. The optimal model uses four convolutional layers followed by fully connected layers, a softmax output layer for three classes (M-CNT, S-CNT, Empty), and cross-entropy loss. Activation maps from the first convolutional layer align with Raman-active CNT features, indicating the network focuses on characteristic spectral bands. Models are trained on datasets collected at specific settings and evaluated on both matched and mismatched acquisition settings to assess generalization. - Generalization tests: Accuracy matrices are produced by training on one integration time (or power) and validating across all integration times (or powers). This assesses robustness to drifts or changes in acquisition conditions. - Softmax thresholding and map classification: For fast Raman maps (e.g., 1 mW, 5 ms), pixels are classified using the CNN. A softmax confidence threshold is applied: pixels are labeled as S-CNT or M-CNT only if the corresponding softmax score exceeds the threshold; otherwise labeled Empty. This reduces false positives while potentially increasing false negatives. Ground truth maps are obtained from long-integration spectra (e.g., 800 ms, 1 mW). Silicon regions (520 cm⁻1) are identified via intensity thresholding separate from the CNN classification.

- High accuracy at low SNR: Despite extremely low SNR (0.9), classification accuracy exceeds 90%; at SNR 2.2, accuracy reaches 98%. - Integration time dependence: Accuracy ~60% at 0.1 ms, >90% at 10 ms, and >98% for integration times >30 ms (at 1 mW). - Power dependence: Accuracy 78% at 0.1 mW (50 ms), peaks at 99% at 1 mW, then slightly decreases to ~96% at higher powers, likely due to CNT heating/damage. - SNR scaling: SNR scales linearly with integration time and laser power, consistent with expectations; SNR < 1 observed for integration times <10 ms at 1 mW and at 0.1 mW with 50 ms. - Generalization behavior: Models trained on low-SNR datasets generalize well to higher-SNR data, but the reverse generalization (high-SNR-trained models to low-SNR data) is poor. For integration times, testing at 20 ms yields >90% accuracy across training settings, and for training/validation times >50 ms, all combinations achieve >98% accuracy. For power, accuracy is highest near matched train/validation settings and degrades as the mismatch increases. - Softmax thresholding: Increasing the softmax confidence threshold in fast maps (e.g., 1 mW, 5 ms) markedly reduces false positives with a smaller reduction in true positives, improving the true-positive/false-positive ratio for both M-CNT and S-CNT classes. Thresholded maps qualitatively reproduce ground truth while suppressing noise-like misclassifications. - Efficient scanning: An optimized line-scan plus rescan strategy localizes CNT positions, counts, and metallicity across forks with reduced acquisition time compared to full mapping.

The study addresses the challenge of high-throughput, reliable identification of metallic vs. semiconducting CNTs under the constraints of low Raman signal and limited acquisition time. By systematically mapping accuracy versus integration time, power, and SNR, the authors establish operating regimes where rapid acquisition still yields high classification accuracy. The observed linear SNR scaling and the demonstrated robustness of models trained on low-SNR data to higher-SNR conditions support deployment in variable experimental environments. The softmax thresholding strategy balances sensitivity and precision in fast maps, curbing false positives that arise from noisy spectra while maintaining the core CNT detections. The optimized scan strategy further reduces measurement time while retaining the ability to localize and quantify CNT occurrences across forks. Together, these elements constitute a practical workflow that can be integrated into fabrication lines for real-time quality control of CNT properties and yields.

The work introduces a high-throughput Raman spectroscopy and deep learning pipeline that rapidly identifies suspended CNTs and distinguishes metallic from semiconducting tubes with high accuracy, even at low SNR. Key contributions include an empirically optimized acquisition regime, a CNN classifier robust to noise and acquisition parameter variations, and a softmax thresholding and rescan strategy for efficient mapping and localization. The approach is generalizable to other nanomaterials and is suitable for integration into production environments to monitor nanomaterial quality and properties. Future directions include expanding training to broader material systems and wavelengths, mitigating laser-induced heating at higher powers, and further optimizing acquisition and thresholding policies for specific deployment constraints.

- Accuracy decreases at very short integration times (e.g., ~60% at 0.1 ms) and at higher powers due to potential CNT heating/damage. - Models trained on high-SNR data generalize poorly to low-SNR conditions, necessitating inclusion of low-SNR data during training. - The mapping approach relies on a specific excitation wavelength; some CNTs may be off-resonance and undetectable, affecting ground truth and classification completeness. - Softmax thresholding reduces false positives but can increase false negatives, potentially missing weak yet real CNT signals. - Reported results pertain to suspended CNTs on fork-like substrates; performance in other geometries/substrates may differ. - Dataset exhibits outliers that can impact certain train/validation combinations; careful curation and augmentation may be needed.