**4.1 Machine learning application in QKDN**

Machine learning (ML) is an application of artificial intelligence (AI) that provides systems the ability to automatically learn and improve from experience without being explicitly programmed. In recent years, a huge amount of attention on ML has been attracted from both the academia and the industry. There has been much development related to ML technologies in both hardware and software. More on-board acceleration chips for neural networks are implemented by new lowpower devices.

Due to the advantages of ML, ML can help to solve several problems in QKDN. In terms of parameter optimization for QKD, ML can greatly improve the efficiency of parameter optimization and allow it to be performed in real time on low-power devices, making it a highly useful tool for both free-space QKD and QKD networks. In terms of key resource utilization, the effectiveness of using reinforcement learning to realize resource allocation in QKD networks is verified, which was published by Asia Communications and Photonics Conference (ACP) 2020 and honored as the best paper award in industry innovation [32]. As for the standardization activities, recommendation ITU-T Y.QKDN-qos-ml-req "Requirements of machine learning based QoS assurance for quantum key distribution networks" specifies the functional mechanisms of machine-learning-based quality of service (QoS) assurance for QKDN; the supplement ITU-T Y.supp.QKDN-mla "ITU-T Y.3800-series - Quantum key distribution networks - Applications of machine learning" specifies different application scenarios of ML in QKDN. In detail, the applications of ML in QKDN include the applications in the quantum layer, key management layer, and QKDN control and management layers of QKDN.


• **The applications of ML in the control and management layers of QKDN** represent applying the ML in the control and management layers and improving the QKDN management and control efficiency: (1) The ML-based data collection and data preprocessing will collect and preprocess multi-source, heterogeneous QKDN data in an efficient way. The collected and preprocessed data will be transformed into understandable, unified, and easy-to-use structures and optimized in the form of balanced characteristics for subsequent procedures. (2) ML-based routing solution will improve the routing effectiveness and the key resources utilization. (3) The ML-based QKDN fault diagnosis solution will reduce the loss and avoid the risk of QKDN faults by realizing fault location and fault prediction.
