Quantum Computing Noise Is Cut with AI-Powered Data Reconstruction

Scientists at City University of Hong Kong, led by Jie Liu and Xin Wang, in collaboration with the Quantum Science Centre of Guangdong-Hong Kong-Macao Greater Bay Area, have developed a novel approach to mitigate the impact of hardware noise on variational quantum algorithms. These algorithms are considered promising candidates for near-term quantum computing applications, but their performance is severely hampered by the inherent errors present in current quantum hardware. The team presents a Physics-Informed Denoising Network (PIDN) designed to accelerate these algorithms by substantially reducing the computational cost associated with error mitigation techniques, specifically Zero-Noise Extrapolation. The method learns to predict the outcomes of noise mitigation, effectively replacing multiple, resource-intensive circuit executions with a streamlined, denoised estimation process. Benchmarking across diverse quantum chemistry and optimisation problems demonstrates that PIDN achieves comparable performance to ZNE while decreasing circuit executions by a factor of four to six, representing a key step towards scalable, resource-efficient quantum computation in the near term.

Physics-informed denoising network accelerates variational quantum optimisation through surrogate modelling

Circuit executions for variational quantum algorithms have been reduced by a factor of approximately four to six using the new Physics-Informed Denoising Network, or PIDN. This decrease in computational cost addresses a critical limitation hindering the scalability of near-term quantum computing. Variational quantum algorithms, such as the Variational Quantum Eigensolver (VQE) and Quantum Approximate Optimisation Algorithm (QAOA), rely on iterative optimisation of parameters within a quantum circuit. However, these algorithms are susceptible to noise arising from imperfections in quantum gates and measurements. Previously, extensive circuit repetitions were necessary for accurate error mitigation via techniques like Zero-Noise Extrapolation. ZNE involves running the quantum circuit at varying levels of artificially introduced noise and then extrapolating to the zero-noise limit to obtain a more accurate result. This process is computationally expensive, requiring a significant number of circuit evaluations. PIDN learns to predict outcomes of this extrapolation, building a ‘surrogate model’ of the optimisation process and maintaining a gradient cosine similarity exceeding 0.95 with ZNE throughout training. This high similarity ensures that the surrogate model accurately captures the direction of optimisation, crucial for achieving reliable results.

Consequently, streamlined, denoised estimation is possible, offering a resource-efficient strategy for noise-resilient optimisation. Benchmarking across multiple quantum algorithms, including those for optimisation and quantum chemistry on systems like LiH and H₂O, revealed performance comparable to Zero-Noise Extrapolation. The LiH (lithium hydride) and H₂O (water) molecules were chosen as representative systems for testing the method’s applicability to quantum chemistry calculations, allowing for comparison with established computational chemistry methods. Removing physics-informed loss functions diminished performance, confirming their necessity in preserving the direction of optimisation. The physics-informed component of the loss function incorporates prior knowledge about the expected behaviour of the optimisation landscape, guiding the network towards solutions that are physically plausible and consistent with the underlying problem. This constraint is vital for preventing the network from learning spurious correlations or diverging towards incorrect solutions.

Its value remains significant despite PIDN’s current accuracy being contingent on the underlying optimisation landscape possessing strong low-frequency structure. A predictable landscape allows for more accurate surrogate modelling, enhancing the efficiency of the quantum computation. Low-frequency structure refers to the presence of smooth, gradual variations in the optimisation surface, as opposed to sharp, erratic fluctuations. In essence, the method performs best when the energy landscape is relatively ‘gentle’. The effectiveness of this approach hinges on the ‘smoothness’ of the problem being solved, as highly complex optimisation landscapes with many twists and turns prove more challenging. This limitation arises because the surrogate model relies on interpolating between known data points, and its accuracy decreases when the underlying function is highly non-linear or contains rapid changes. A more rugged landscape requires a more complex model to accurately capture the variations, potentially negating the computational benefits of the PIDN approach.

PIDN accelerates calculations for a range of quantum algorithms, which is important given the current limitations in quantum computer scale and error rates. The number of qubits and the fidelity of quantum gates are currently major bottlenecks in quantum computing. Reducing the number of circuit executions required for error mitigation can significantly alleviate these limitations, enabling the exploration of larger and more complex problems. Future work should investigate adapting it to handle more rugged problem spaces, acknowledging that its current reliance on a predictable landscape represents a limitation. Exploring alternative network architectures or loss functions designed to better capture high-frequency variations in the optimisation surface could address this challenge. This might involve incorporating techniques from signal processing or machine learning to enhance the network’s ability to model complex functions.

Accelerating quantum computation through learned error correction prediction

Current quantum devices suffer from inherent errors that plague even the most promising algorithms. These errors arise from various sources, including decoherence, gate infidelity, and measurement errors. Researchers in Quantum AI are striving to build practical quantum computers capable of solving problems intractable for classical computers. A new Physics-Informed Denoising Network offers a clever shortcut by learning to predict the results of complex error correction, reducing the need for repeated calculations. Error mitigation techniques, such as ZNE, aim to reduce the impact of these errors without requiring full-fledged quantum error correction, which is still beyond the capabilities of current hardware. A reduction of four to six times in repeated quantum circuit executions represents a sharp step towards scalable quantum computation, particularly for optimisation and quantum chemistry problems. The method’s performance was assessed by comparing its results to those obtained using Zero-Noise Extrapolation, a well-established error mitigation technique. This comparison provides a benchmark for evaluating the accuracy and efficiency of the PIDN approach.

The research demonstrated a Physics-Informed Denoising Network successfully predicted the performance of Zero-Noise Extrapolation, a technique used to lessen errors in quantum computations. This matters because it significantly reduces the number of times quantum circuits need to be run, by a factor of approximately 4 to 6, without compromising accuracy. The network learned to reproduce expectation values and gradient directions, preserving the essential dynamics of the optimisation process. Authors suggest future work will focus on adapting the method to more complex computational landscapes.