应用机器学习方法对量子点自旋量子比特抗噪调控的理论研究

Quantum dot spin qubits are among the most promising candidates for the physical realization of quantum computing. The key problem to accurately control them is to minimize influences from environmental noises. The applications of dynamically corrected gates based on traditional theoretical methods are limited, while the development of artificial intelligence sheds new light to the problem. In this project, we use the new methodology of machine learning to investigate and solve problems pertaining to noise-resistant control from a new angle, including: First, based on supervised learning we improve the efficiency of quantum gate design. Further, we use randomized benchmarking to provide a viable method to measure the noise spectra in quantum-dot devices. Next, based on reinforcement learning and the deep-Q network, we investigate the quantum state preparation and transfer, as well as the development of the next-generation quantum gates. In the quantum state transfer, we expect reinforcement learning can break through known quantum speed limits; for the next-generation quantum gates, we encapsulate into the rules of the reinforcement learning the physical requirements of optimizing the quantum gates, including the complexity of the pulse sequences and the sensitivity to certain specific noise spectra, thereby let the machine automatically find the optimal solution. This "filter-function engineering" approach shall be immensely useful for practical applications. We expect 10-15 scientific journal papers as part of the outcome of this project.

量子点自旋量子比特作为实现量子计算的候选体系之一，对其精确调控的核心在于如何使环境噪声的影响最小化。传统的动力学修正量子门的效能较为有限，而人工智能的发展为我们提供了全新的思路。在本项目中，我们利用机器学习的新方法以全新角度审视并尝试解决抗噪操控所衍生出的若干问题，包括：首先，基于监督学习提高设计量子门的效率，并进一步以随机标定技术给出测量量子点器件的噪声谱的一个可行方法。其次，基于强化学习及深度价值网络，研究量子态制备、转移及下一代抗噪量子逻辑门的设计。在量子态转移中，我们期待强化学习能够突破已知的量子速度极限；而对于下一代抗噪量子门，我们将实际问题所需要优化的物理内容例如脉冲序列的复杂性，尤其是对特定频谱噪声的敏感性设计进强化学习的规则当中去，从而使得机器自动找到问题的最优解。这种按实际需要而设计过滤函数的方法将有潜力得到广泛的应用。本项目预期发表学术论文10～15篇。

在该项目中，我们给出了具有非正交旋转轴的单量子比特幺正量子门优化分解问题的完整解决方案；研究并展示了一种通过监督学习辅助从随机标定过程中获取量子器件中特定噪声频谱的方法；研究了量子态制备和自旋态在一维链传输的问题，揭示了强化学习算法的自适应性和高效率；提出了一个基于检测的量子自动编码器来减少量子控制的误差，其对于近期量子器件效果显著；通过将强化学习应用于量子参数估计的研究揭示了强化学习在量子控制的应用中的可拓展性，使得其在对系综进行大规模测量时与传统算法相比拥有极大的时间优势。本项目共发表论文15篇，包括npj Quantum Information 2篇, Physical Review A 7篇，Physical Review B 3篇, Advanced Quantum Technologies 3篇。本项目的研究目标圆满完成。