Efficient tuning of quantum devices using machine learning

拥有大量与技术相关的量子比特的量子计算机似乎有潜力在从传感器技术到高性能计算等领域取得突破性进展。有几个相互竞争的硬件候选者可以实现这些大型量子计算机。硅器件已经展示了对电子自旋的高保真控制，并实现了长松弛和去相时间，这使它们成为最受欢迎的可扩展量子比特候选者。可扩展性方面的关键挑战之一是基于硅的更大量子系统的校准和控制；使这一点变得困难的是设备的可变性和热循环后行为的变化。即使对于半导体器件来说，在最小化变化方面取得了巨大的进步，缺陷仍然存在，很难估计。随着这些技术的规模扩大，传统的表征方法变得难以处理，这突显了对高效自动化的迫切需要。为了解决这一问题，我们建议开发一种实时贝叶斯机器学习算法，将信息论与概率深度生成模型相结合。这种创新的组合将使算法能够智能地选择下一步执行的信息最丰富的测量，从而简化表征过程。初步结果表明，我们的方法明显优于标准的网格扫描方法，减少了测量次数和表征所需的时间。重要的是，我们的工作超越了传统的使用机器学习进行数据搜索和分析，而是展示了它在自动化设备表征所需的实际测量方面的有效性。通过实施这种贝叶斯方法，我们的目标是为基于学习的自动化测量技术奠定基础，为大规模、高效地部署量子技术打开大门。由于在真实量子设备上进行测量的成本很高，用于训练机器学习模型的数据量是有限的。我们将研究如何将量子主方程或恒定电容模型等物理模型融入到机器学习方法中。该项目是释放量子技术潜力的关键，它允许快速调整量子设备，从而有可能控制技术上相关的量子电路。该项目属于EPSRC量子技术研究领域。这项工作将作为牛津大学工程系物理学家和机器学习科学家之间的合作进行。此外，我们将与埃尔兰根的Max-Plank光科学研究所合作。

Quantum computers with a technologically relevant number of qubits seem to have the potential for ground-breaking advances in areas ranging from sensor technology to high-performance computing. There exist several competing hardware candidates to realise these large quantum computers. Silicon devices have demonstrated high-fidelity control of electron spins and achieved long relaxation and dephasing times, which make them a favourite scalable qubit candidate. One of the key challenges in scalability is the calibration and control of larger quantum systems based on Silicon; what makes this difficult is device variability and changed behaviour after thermal cycling. Even for semiconductor devices where huge progress was achieved in minimizing variations, defects are still present and difficult to estimate. As the scale of these technologies expands, traditional methods of characterization become intractable, underscoring the urgent need for efficient automation. To address this, we propose to develop a real-time Bayesian machine learning algorithm that integrates information theory with probabilistic deep-generative models. This innovative combination will enable the algorithm to intelligently select the most informative measurements to perform next, thereby streamlining the characterization process. Preliminary results indicate that our approach significantly outperforms standard grid scan methods, reducing both the number of measurements and the time required for characterization. Importantly, our work goes beyond the conventional use of machine learning for data search and analysis, and instead demonstrates its utility in automating the actual measurements necessary for device characterization. By implementing this Bayesian approach, we aim to lay the foundation for learning-based automated measurement techniques, opening the door for the large-scale, efficient deployment of quantum technologies. Due to the high cost of measuring on real quantum devices the amount of data to train machine learning models is limited. We will investigate ways on how to incorporate physical models like quantum master equation or the constant capacitance model into the machine learning approach. This project is key to unleashing the potential of quantum technologies by allowing fast tuning of quantum devices, and thus the possibility to control technologically relevant quantum circuits. This project falls within the EPSRC Quantum technologies research area. This work will take place as a cooperation between physicists and machine learning scientists within the Engineering Department of the University of Oxford. Furthermore, we will collaborate with the Max-Plank Institute for the Science of Light in Erlangen.