Machine learning with neural networks for parameter optimization in twin-field quantum key distribution

Twin-field quantum key distribution (TF-QKD) has the advantage of beating the rate-loss limit (PLOB bound) for a repeaterless quantum key distribution (QKD) system. In practice, parameter optimization is of great significance in maximizing the secret key rate. Nevertheless, traditional local search algorithms (LSA) are often time-consuming and limited by the computing capabilities of devices. In this paper, we use the machine learning method instead of LSA to directly predict the optimal parameters for TF-QKD system. Specifically, three neural networks, namely back propagation neural network, radial basis function neural network, and generalized regression neural network, are trained and evaluated. The performance of neural networks and LSA in optimizing parameters is discussed and analyzed in this study. It is proved that the performance of machine learning-based prediction method is comparable to LSA, but the calculation time is shortened by 6 orders of magnitude. Furthermore, a comprehensive comparison of three networks in terms of prediction accuracy and time consumption is conducted, serving as a guide for selecting the most suitable network to optimize parameters in a practical TF-QKD system with different optimization requirements.