<p>Quantum key distribution (QKD) enables information-theoretically secure communication, even in the era of quantum information. In all QKD systems, clock synchronization between two remote users—commonly referred to as Alice and Bob—is a fundamental requirement. This is typically achieved by transmitting an additional reference clock signal from Alice to Bob. In such a scheme, additional synchronization devices are required, increasing system complexity and introducing external noise. To address these issues, a novel synchronization technology, called the qubit-based synchronization method, was proposed. This method directly synchronizes two users using quantum signals, thereby dramatically reducing system complexity. However, previous qubit-based synchronization methods are not applicable to time-bin phase-encoding QKD systems, as multiple time slides introduce disturbances to time recovery. In this paper, we propose a machine-learning-enhanced qubit-based synchronization method. By introducing a K-nearest neighbor model, this method can efficiently classify each time slide in time-bin phase-encoding QKD, thereby enabling successful time recovery. We demonstrate our method using a time-bin phase-encoding reference-frame-independent (RFI)-QKD and successfully distribute secure key bits over up to 200 km of fiber spools. Our work simplifies the complexity of QKD system and significantly advances the practical application of QKD.</p>