Machine Learning Extends Quantum Data Processing Beyond Current Limits

A new quantum error mitigation (QEM) framework corrects for noise in continuous-variable (CV) quantum systems, addressing a key challenge that degrades quantum states vital for quantum information processing. Jingpeng Zhang and colleagues from the School of Physics, Sun Yat-sen University, in collaboration with Qianchuan Zhao from the Department of Automation, Tsinghua University, and Jie Han also at Sun Yat-sen University and the State Key Laboratory of Optoelectronic Materials and Technologies, developed this framework. Continuous-variable quantum systems utilise infinite-dimensional Hilbert spaces, offering advantages in encoding and manipulating quantum information compared to qubit-based systems, but are particularly susceptible to environmental noise that limits their coherence and fidelity. This noise arises from interactions with the surrounding environment, leading to photon loss and dephasing, which progressively corrupt the quantum state.

Their research introduces an extrapolative approach, using a time-conditioned Swin Transformer to correct for noise accumulation without needing extensive training data covering the entire quantum evolution. This advancement represents a sharp step towards practical noise mitigation in CV quantum systems, demonstrating accurate state recovery even in long-time regimes where conventional methods fail. Traditional QEM techniques often rely on characterising the noise and applying complex error correction codes, demanding significant computational resources and precise knowledge of the noise process. The ability to mitigate errors with limited training data is therefore crucial for scaling up CV quantum systems and realising their full potential.

Time-conditioned machine learning corrects quantum states beyond training data

Error rates dropped to 0.6%, a substantial improvement over existing methods which typically fail beyond their initial training parameters. This breakthrough enables accurate quantum state recovery even when extrapolating beyond the data used to train the system, something previously unattainable in continuous-variable (CV) quantum computing. Naren Manjunath from the Perimeter Institute and colleagues at Sun Yat-sen University and Tsinghua University developed a time-conditioned Swin Transformer, a machine learning model that explicitly models how noise accumulates over time. Swin Transformers, originally developed for computer vision, excel at capturing long-range dependencies in data, making them well-suited for modelling the complex correlations present in quantum phase space. The ‘time-conditioned’ aspect allows the model to incorporate the temporal evolution of the quantum state, crucial for accurately predicting and correcting noise accumulation.

The model captures subtle, long-range correlations within the quantum system’s phase space, allowing for more reliable error correction. Adaptive layer normalization proved key to modelling this accumulation, enhancing reliability. Layer normalization is a technique used to stabilise the training process of deep neural networks, and its adaptive implementation allows the model to dynamically adjust its internal parameters based on the time evolution of the quantum state. Quantum states were successfully recovered even with non-Markovian noise, a complex form of environmental disruption. Markovian noise assumes that the future state depends only on the present state, while non-Markovian noise incorporates memory effects, making it significantly more challenging to mitigate. Simulations revealed the framework’s ability to accurately reconstruct states at timescales exceeding those used for initial training, a key limitation of previous machine learning-based quantum error mitigation techniques. This extrapolation capability is vital for performing longer and more complex quantum computations. While the framework achieved an impressive 0.6% error rate in simulations, these results currently rely on idealised conditions and do not yet demonstrate performance with real, imperfect quantum hardware.

Time-dependent error mitigation via adaptive normalisation of Swin Transformers

A time-conditioned Swin Transformer forms the core of this advancement, a machine learning model employed to decipher patterns in complex data. Unlike previous methods, this model doesn’t demand exhaustive training data covering the entire quantum process. Instead, it learns to correct errors by explicitly accounting for how noise accumulates over time, embedding the evolution time using adaptive layer normalization. The Swin Transformer architecture utilises a hierarchical structure with shifted windows, enabling efficient processing of high-dimensional data while capturing both local and global features. By conditioning the model on time, the framework can effectively learn the dynamics of noise and predict its impact on the quantum state at different points in time.

This allows the model to build a ‘correction map’ that anticipates and mitigates the effects of noise even beyond the initial training period. Numerical simulations were performed on continuous-variable quantum systems, utilising a qubit count equivalent to a 10-dimensional phase space. These systems experienced both photon loss and dephasing, common sources of environmental noise, at varying intensities. Photon loss represents the irreversible loss of photons from the quantum system, while dephasing refers to the loss of phase coherence, both contributing to the degradation of quantum information. The 10-dimensional phase space represents the range of possible quantum states, allowing for a comprehensive assessment of the framework’s performance.

The simulations employed both Markovian and non-Markovian noise models to assess the framework’s strong resistance against different types of environmental disruption. Previous methods required complete evolution data, a significant experimental hurdle. The team opted for a machine learning approach to quantum error mitigation, avoiding exhaustive data collection by focusing on time-dependent error correction. This reduces the burden of extensive calibration required by existing methods, a major obstacle to scaling up quantum computers. Traditional calibration methods involve characterising the noise at every point in time, which is computationally expensive and time-consuming. By learning the noise dynamics, the framework can significantly reduce the amount of calibration data required.

Reducing data demands improves durability in continuous-variable quantum computation

A new quantum error mitigation framework has been engineered, sidestepping the need for exhaustive data collection in continuous-variable (CV) quantum systems – a vital step towards building more powerful and reliable quantum computers. However, the current approach relies on simulations, raising questions about its performance with real-world quantum hardware susceptible to unpredictable imperfections. The transition from simulation to real hardware presents significant challenges, as real devices are subject to imperfections in fabrication, control, and measurement, which are not fully captured in simulations.

It is important to acknowledge that this framework currently relies on simulations rather than direct testing with physical quantum hardware. Despite this limitation, the development of a quantum error mitigation technique requiring sharply less data represents a substantial advance. The team’s simulations explored the framework’s behaviour with varying noise intensities, providing insights into its durability under different conditions. Understanding the framework’s robustness to different noise levels is crucial for optimising its performance in real-world scenarios.

This detailed analysis will be crucial for adapting the framework to the complexities of real quantum devices. By employing a time-conditioned Swin Transformer, the framework accurately corrects for noise accumulating during quantum calculations, capturing subtle correlations within the quantum system. No prior method matched this. A new approach to quantum error mitigation for continuous-variable systems has been established, bypassing the need for extensive training data. Future work will focus on implementing this framework on physical CV quantum systems and evaluating its performance in realistic experimental settings, paving the way for more robust and scalable quantum computation.