Not yet reviewed by Pith; the record is open.

Despite the robust security guarantees of Quantum Key Distribution (QKD), practical deployment is hindered by dynamic channel noise and complex parameter optimization. We propose AML-QKD, a protocol-agnostic machine learning framework designed to maximize the Secure Key Rate (SKR) and minimize the Quantum Bit Error Rate (QBER) across the BB84, E91, and COW protocols. AML-QKD integrates Temporal Convolutional Networks (TCNs) for short-horizon forecasting of channel fluctuations, using a Proximal Policy Optimization (PPO) agent for real-time parameter tuning, while strictly adhering to composable security constraints. Simulations under realistic depolarizing and amplitude-damping noise demonstrate a 14-25% increase in median SKR and a reduction in median QBER from 3.0% to 1.5%. Furthermore, an exploratory Quantum Reinforcement Learning (QRL) extension reveals a distinct quantum advantage for entanglement-based protocols (E91), achieving a 29.2% throughput gain by natively processing non-local correlations. Our findings suggest that AML-QKD can offer a potentially resilient, security-preserving control architecture for next-generation quantum networks.