Deep learning for classifying quantum emission signals in WS2 monolayers using wavelet transform

<jats:title>Abstract</jats:title> <jats:p> This study aimed to develop and evaluate deep learning approaches for the classification of quantum emission signals from WS <jats:sub>2</jats:sub> monolayer nanobubbles across multiple spectral bands, addressing challenges in quantum materials characterization and spectral distinguishability assessment. We utilized a dataset of quantum emission signals ranging from 604 to 629 nm, emitted from WS₂ monolayer nanobubbles on gold substrates, categorized into four spectral bands (604.06–608.24 nm, 611.07–616.34 nm, 617.42–623.35 nm, and 624.16–636.57 nm). Our methodology involved signal preprocessing through normalization and moving average smoothing, followed by transformation into 128 × 128 RGB images using Continuous Wavelet Transform (CWT) with Complex Morlet wavelet. Three convolutional neural network architectures (ResNet50, VGG16, and Xception) were implemented and evaluated using fivefold cross-validation across six possible band pair combinations. All models demonstrated exceptional classification performance, with VGG16 achieving the highest overall mean accuracy of 99.4%, followed by Xception (99.1%) and ResNet50 (98.2%). Perfect classification accuracy (100%) was consistently achieved for spectrally distant band pairs, particularly Band 1 versus Band 4 (20.5 nm separation), while the most challenging classification involved adjacent bands (Band 2 vs. Band 3, 6.27 nm separation) with VGG16 achieving 96.5% accuracy. Statistical analysis using Friedman tests confirmed significant performance differences among models (χ <jats:sup>2</jats:sup>  = 8.67, <jats:italic>p</jats:italic>  = 0.013). Xception demonstrated remarkable computational efficiency, achieving optimal convergence in as few as 2 epochs for certain band combinations while maintaining ultralow training loss values (8.23 × 10⁻ <jats:sup>6</jats:sup> ). Deep learning models, particularly when combined with CWT preprocessing, provide a robust framework for quantum emission signal classification with significant implications for quantum photonics, quantum cryptography, and quantum sensing applications. Our approach bridges the gap between classical machine learning and quantum materials characterization, establishing quantifiable metrics for evaluating spectral distinguishability in quantum information systems. The demonstrated ability to achieve high classification accuracy with minimal training through transfer learning addresses data scarcity challenges inherent to quantum systems, offering a promising direction for future quantum technology development. </jats:p>