[Paper] Domain-Aware Quantum Circuit for QML

Source: arXiv - 2512.17800v1

Overview

The paper introduces Domain‑Aware Quantum Circuit (DAQC), a new way to build parameterized quantum circuits (PQCs) for image‑classification tasks on noisy intermediate‑scale quantum (NISQ) hardware. By weaving image‑specific priors (e.g., locality of pixels) into the circuit layout, DAQC achieves accuracy that rivals heavyweight classical models while staying within the tight depth and qubit budgets of today’s quantum processors.

Key Contributions

• Locality‑preserving encoding: Uses a DCT‑style zig‑zag window to map neighboring pixels onto physically adjacent qubits, respecting the device’s connectivity graph.
• Interleaved encode‑entangle‑train cycles: Alternates between data encoding, short‑range entanglement, and parameter updates, avoiding deep global mixing that typically causes barren‑plateau problems.
• Hardware‑aware entanglement pattern: Reduces the number of long‑range two‑qubit gates, which are the primary source of noise on NISQ devices.
• Empirical benchmark: Demonstrates state‑of‑the‑art performance on real quantum hardware for MNIST, FashionMNIST, and PneumoniaMNIST, beating prior Quantum Circuit Search (QCS) baselines and matching strong classical CNNs (ResNet‑18/50, DenseNet‑121, EfficientNet‑B0).
• Open‑source release: Provides code, pretrained models, and a reproducible training pipeline (GitHub link in the abstract).

Methodology

1. Image‑to‑Qubit Mapping – The input image is first transformed with a discrete cosine transform (DCT). Pixels are then grouped into non‑overlapping “zig‑zag” windows that preserve spatial locality. Each window is assigned to a small block of qubits that are physically adjacent on the quantum chip.
2. Encode‑Entangle‑Train Loop –
   • Encode: Parameterized rotation gates (R_X, R_Y, R_Z) embed the pixel values into the qubit amplitudes.
   • Entangle: A shallow layer of CNOT (or native CX) gates is applied only between qubits that host neighboring pixels, mirroring the image’s local correlation structure.
   • Train: The circuit parameters are updated via gradient‑based optimizers (e.g., Adam) using the parameter‑shift rule, while a simple linear readout layer maps measurement outcomes to class scores.
3. Depth Management – Because each entanglement layer touches only short‑range qubit pairs, the overall circuit depth stays low (typically ≤ 15 layers), which keeps decoherence and gate‑error rates manageable.
4. Hardware Alignment – The entanglement pattern is explicitly matched to the connectivity map of the target quantum processor (e.g., IBM’s heavy‑hex lattice), avoiding SWAP overhead.

Results & Findings

• Depth vs. performance: Adding more global entanglement layers beyond the locality‑preserving ones yields diminishing returns and quickly triggers barren‑plateau behavior.
• Noise robustness: The reduced count of long‑range two‑qubit gates cuts the overall error per circuit by ~30 % compared with a naïve, fully‑connected PQC of similar size.
• Training stability: Loss curves show smoother convergence and lower variance across random seeds, indicating that the domain‑aware layout mitigates the “vanishing gradient” problem common in deep PQCs.

Practical Implications

• Ready‑to‑use quantum feature extractor: Developers can plug the DAQC encoder into existing hybrid pipelines (e.g., combine with a classical transformer or a downstream decision tree) without needing a deep classical backbone.
• Hardware‑conscious design pattern: The paper provides a template for aligning algorithmic structure with device topology—a practice that can be generalized to other QML tasks such as time‑series forecasting or graph classification.
• Reduced compilation overhead: By matching the entanglement pattern to native connectivity, the need for costly SWAP insertion is eliminated, shortening compile times and freeing up qubit resources for larger datasets.
• Benchmark for NISQ‑level AI: DAQC sets a new performance baseline for image classification on real quantum hardware, giving product teams a concrete target when evaluating the ROI of quantum‑accelerated inference in edge or privacy‑sensitive scenarios.

Limitations & Future Work

• Scalability to high‑resolution images: The current approach relies on a fixed number of qubits (≈ 27); scaling to larger images would require clever patching or hierarchical encoding strategies.
• Hardware dependence: The entanglement pattern is tuned to IBM’s heavy‑hex lattice; porting to other architectures (e.g., trapped‑ion chains) may need redesign of the locality windows.
• Classical readout simplicity: While a linear readout keeps the quantum part lightweight, more expressive classical post‑processing could further close the gap with deep CNNs.
• Barren‑plateau analysis: The authors note that deeper global mixing still leads to barren plateaus; future work could explore adaptive depth schedules or noise‑aware cost functions to push beyond the current depth ceiling.

The authors have released their implementation and pretrained models, making it easy for developers to experiment with DAQC on their own quantum hardware or simulators.

Authors

• Gurinder Singh
• Thaddeus Pellegrini
• Kenneth M. Merz

Paper Information

• arXiv ID: 2512.17800v1
• Categories: quant-ph, cs.LG
• Published: December 19, 2025
• PDF: Download PDF