Introduction
Semiconductor manufacturing is a high-stakes, resource-intensive process where undetected defects lead to significant waste in energy, water, and raw materials. While classical Deep Learning (DL) has improved defect review, it faces computational bottlenecks—the "curse of dimensionality"—when processing complex, high-dimensional data. This paper introduces HCQDL (Hybrid Classical-Quantum Deep Learning), a framework that integrates a novel classical architecture with Parametrized Quantum Circuits (PQC) to achieve state-of-the-art accuracy in defect detection.
The HCQDL Framework
The proposed architecture bridges the gap between classical efficiency and quantum expressivity through three main components:
1. Self-Proliferation-and-Attention (SP&A) Block
The classical front-end uses a custom SP&A block designed for efficiency.
- Self-Proliferation: Inspired by DNA replication, this mechanism uses a small set of convolutions followed by inexpensive linear transformations. This generates a large number of feature maps at a fraction of the computational cost of standard CNNs.
- Self-Attention: This component captures long-range dependencies across the wafer image, providing the context necessary to distinguish between different defect patterns, such as identifying whether a particle is part of a "cluster" or an "edge" defect.
2. Parametrized Quantum Circuit (PQC)
Replacing traditional fully connected linear layers, the PQC serves as a non-linear booster. Features are encoded into quantum states using methods like Basis, Amplitude, or Angle encoding. The authors tested various circuit topologies, finding that "fully connected" quantum graphs (such as Circuits 5 and 6) provided the best expressibility and entangling capability. These circuits utilize rotation gates ($R_x, R_y, R_z$) and CNOT gates to entangle qubits, creating complex kernels that exploit quantum superposition to identify patterns in high-dimensional space.
3. Classification Layer
A final classical fully connected layer with a Softmax activation function produces the classification probability for the specific defect type.
Experimental Results
The HCQDL model was evaluated on two major industry benchmarks: ICCAD-2012 for lithography hotspot detection and WM-811K for wafer map defect pattern classification.
The hybrid approach consistently outperformed pure classical models and previous state-of-the-art (SOTA) methods:
- Lithography Hotspots (ICCAD-2012): Achieved 99.12% accuracy, surpassing the previous SOTA of 99.01%.
- Wafer Maps (WM-811K): Reached 98.10% accuracy, a significant improvement over the previous best of 97.30%.
Ablation studies confirmed that replacing classical layers with even small-scale quantum circuits (4 qubits) provided a consistent performance lift across various backbones, including ResNet50, SENet, and MobileNetV3.
Conclusion and Future Directions
The success of HCQDL demonstrates that quantum circuits can effectively replace linear portions of traditional neurons to find complex patterns. However, the study acknowledges the limitations of the current "Noisy Intermediate-Scale Quantum" (NISQ) era. Future research will focus on automated circuit selection and moving from simulations to actual near-term quantum hardware to quantify real-world speedups. This work sets a roadmap for achieving "quantum advantage" in semiconductor manufacturing within the next decade.
Summary updated