Physics-Informed Machine Learning for Route Feasibility Prediction in Quantum Networks

Quantum networking promises transformative advances in secure communication, yet entanglement distribution remains fragile due to photon loss, decoherence, and memory constraints, making analytical route selection impractical for dynamic operations. While machine learning has been explored for quantum routing, existing approaches often lack systematic generalization testing on unseen topologies, physics-based feature engineering, and rigorous cross-validation. This work proposes a physicsinformed machine learning framework for route feasibility prediction and routing policy selection. Using the SeQUeNCe simulator, we generate datasets from an eight-router asymmetric mesh for development and an unseen irregular ladder for external evaluation. Eleven physics-informed prerouting features capture fibre loss, hop count, bottleneck success, memory margins, and coherence time. A nested grouped cross-validation strategy (outer 5-fold, inner 4-fold) with scenario-level stratification mitigates data leakage and ensures robust model selection. Logistic regression is selected for interpretability and strong performance, with the operating threshold locked at 0.405 on validation data. Evaluation on the unseen ladder topology achieves a ROCAUC of 0.950 and an F1-score of 0.905, demonstrating a reliable cross-topology generalization. As a routing policy, the learned model reduces mean route-selection regret by 50.0% relative to random selection and remains competitive with deterministic heuristics. The framework establishes a
rigorous, reproducible protocol for data-driven quantum routing, highlighting the viability of interpretable models for operational network management.