Yonsei University Team Builds Continuous-Variable QNN for Wafer-Map Defect Classification

A new study directly compares continuous-variable (CV) and discrete-variable (DV) quantum neural networks for classifying defects on WM-811K wafer maps, a key step in semiconductor yield optimisation. Yeonhong Kim at Yonsei University, and colleagues in collaboration with Vellore Institute of Technology, employed a shared convolutional network with interchangeable quantum heads scaled from three to eight qumodes/qubits. The results reveal a consistent performance advantage for CV quantum neural networks, achieving 79.7 ±1.8% accuracy compared to 61.6 ±1.4% for DV systems. This 18-point gap is particularly pronounced in distinguishing spatially localised defects, highlighting the potential of structured CV layers and continuous phase-space encoding to capture subtle defect characteristics. This offers a pathway towards practical quantum advantage as noise and scale improve.

Continuous-variable quantum neural networks demonstrate superior silicon wafer defect classification

Accuracy reached 79.7 ±1.8% using continuous-variable quantum neural networks (CV-QNNs), exceeding the performance of discrete-variable quantum neural networks (DV-QNNs) by 18 percentage points in classifying silicon wafer defects. This represents a key step towards practical quantum advantage in industrial applications. Discerning subtle spatial differences in wafer defects previously proved challenging for quantum neural networks, limiting their potential for yield optimisation in semiconductor manufacturing. The increasing complexity of integrated circuits necessitates increasingly sensitive defect detection methods, as even minor imperfections can significantly impact device performance and reliability. Traditional inspection techniques, while effective for larger defects, often struggle with the nanoscale flaws becoming prevalent in advanced semiconductor fabrication processes.

The CV-QNN’s success demonstrates an ability to capture these fine distinctions, particularly with spatially localised defects like Edge-Loc. A dataset of 2,552 wafer maps, categorising eight defect classes, was used to assess performance. These defect classes represent a range of common issues encountered during wafer fabrication, including scratches, edge-linkage defects, and various forms of contamination. The use of the WM-811K wafer map dataset is significant as it is a standard benchmark within the semiconductor industry, allowing for direct comparison with existing classical methods. The CV-QNN exhibited a sharply higher recall of 0.66 ±0.06 for spatially localised Edge-Loc defects, which the DV-QNN frequently misidentified as Scratch defects and failed to correctly identify at any scale. This misclassification highlights a fundamental limitation of the DV approach in representing the nuanced spatial features of these defects. Training dynamics revealed that the DV-QNN’s limitations stem from a representational capacity ceiling, unable to effectively fit even its training data, while the CV-QNN reached 99% training accuracy. This suggests that the continuous nature of the CV encoding allows for a richer and more flexible representation of the input data, enabling the network to learn more complex patterns. Although both quantum approaches currently fall short of the 85.0% accuracy achieved by classical convolutional neural networks, these results pinpoint where a structured quantum head offers benefit, but do not yet demonstrate practical advantage given current noise levels and scale.

Quantum neural network performance for semiconductor quality control

The relentless drive for smaller, faster chips fuels a constant need to refine wafer defect screening; identifying flaws during manufacturing is vital, especially as advanced components for artificial intelligence and memory become increasingly complex. A team from Yonsei University and the Vellore Institute of Technology has shown that continuous-variable quantum neural networks show promise in this area, outperforming discrete-variable counterparts. Their detailed comparison of continuous-variable and discrete-variable quantum neural networks clarifies which approach better suits wafer defect classification, with the structured layers within continuous-variable systems adept at discerning subtle spatial differences in flaws. The motivation behind exploring quantum machine learning for this task stems from the potential for exponential speedups in pattern recognition, which could significantly reduce inspection times and improve overall manufacturing efficiency.

The core difference between CV and DV quantum computing lies in the information carriers. DV qubits utilise discrete energy levels to represent information, analogous to classical bits, while CV qubits encode information in continuous degrees of freedom, such as the amplitude and phase of light. This continuous encoding allows for a more natural representation of complex data, potentially leading to improved performance in tasks like image classification. The researchers employed a shared convolutional network architecture, meaning that both the CV-QNN and DV-QNN used the same classical convolutional layers for feature extraction. This ensured that any performance differences could be directly attributed to the quantum heads, isolating the impact of the quantum paradigm. The quantum heads were scaled from three to eight qumodes/qubits, allowing the researchers to investigate the effect of quantum resource allocation on performance. The use of eight qumodes/qubits represents a significant step towards scaling up quantum neural networks for more complex tasks.

Further refinement is anticipated to unlock practical applications in the coming decade. The findings from Yonsei University and the Vellore Institute of Technology establish a performance distinction between continuous-variable and discrete-variable quantum neural networks when applied to wafer-map defect classification, a vital process for semiconductor manufacturing. The continuous-variable approach consistently outperformed its discrete-variable counterpart, particularly in identifying spatially localised defects like Edge-Loc, demonstrating an ability to capture subtle differences important for yield optimisation. While current quantum hardware limitations prevent immediate deployment, ongoing advancements in qubit coherence and gate fidelity are expected to pave the way for practical quantum-enhanced wafer inspection systems. Future research will focus on mitigating the effects of noise and exploring more sophisticated quantum network architectures to further improve performance and scalability. The potential impact on semiconductor manufacturing is substantial, promising increased yields, reduced costs, and ultimately, more powerful and reliable electronic devices.

The research demonstrated that continuous-variable quantum neural networks achieved 79.7 +/- 1.8% accuracy in classifying wafer-map defects, outperforming discrete-variable networks which reached 61.6 +/- 1.4%. This difference in performance is significant because it suggests continuous-variable approaches are better at recognising subtle spatial patterns within defect types, crucial for optimising semiconductor manufacturing yields. Researchers used a shared convolutional network with varying numbers of qumodes/qubits to isolate the impact of each quantum paradigm. The authors intend to address noise and explore improved network architectures to further enhance performance and scalability.