Genetic algorithm-enhanced neural network for parameter optimization of quantum key distribution

Quantum Key Distribution (QKD) parameter optimization is crucial for maximizing secure key rates under realistic finite-key conditions. However, the commonly used Local Search Algorithm (LSA) exhibits clear limitations in protocols with non-convex objective functions or extremely narrow feasible regions. This paper proposes a neural-network framework enhanced by a Genetic Algorithm (GA) for QKD parameter optimization as a practical alternative to LSA-based data generation. We develop a customized GA for QKD optimization to construct higher-quality training datasets, and combine it with a multi-head residual neural network to predict both near-optimal key rates and corresponding protocol parameters. Experimental validation shows that, for symmetric and asymmetric Bennett–Brassard 1984 (BB84) protocol, the GA-based models achieve average parameter-prediction improvements of 14.52% and 42.03%, respectively, over the corresponding LSA-based models, where these improvements are computed over the controllable protocol parameters using the regression metrics reported in the main text. In addition, we evaluate the predicted parameters by substituting them back into the original QKD key-rate formula, thereby obtaining physically meaningful performance comparisons that do not rely solely on optimizer-generated targets. Furthermore, we extend the framework to the more complex Reference-Frame-Independent Measurement-Device-Independent QKD (RFI-MDI-QKD) protocol. The results show that our method remains effective in this setting, whereas LSA-based optimization performs poorly for several parameters.