Energy-efficient ML-based resource allocation in 6G x-haul networks utilizing a photonic interleaver-based subcarrier switching node

We investigate energy-efficient, traffic-adaptive resource provisioning for 6G x-haul networks by integrating a reconfigurable low-loss switching sub-system (referred to as an interlacer) into the TEFNET24 reference network topology from Telefónica. The interlacer combined with the digital subcarrier multiplexing (DSCM) transceivers forms a WDM-PON approach, enabling subcarrier-level switching at the network access, providing lower distribution loss than conventional power splitter (PS)-based TDMA-PON approaches. Moreover, under the practically low crosstalk levels reported for our interlacer implementations, the reduced distribution loss can be traded for increased modulation cardinality (e.g., 16QAM to 64QAM), allowing for reducing the number of allocated transceivers. We further examine a machine learning (ML)-driven framework for dynamic capacity provisioning and validate it using a real-world operator trace mapped onto the national-scale TEFNET24 topology, capturing 4G/5G load patterns with projections toward 6G. We compare static, semi-static, and fully dynamic strategies and show that ML-aided dynamic provisioning reduces transceiver overallocation by up to 63.7% relative to static methods, with the strongest gains during off-peak hours when traffic variability is highest. Finally, we quantify two complementary energy-saving mechanisms enabled by dynamic operation-subcarrier deactivation and transceiver sleep mode-with the latter yielding up to 21% total network energy savings while maintaining quality of service under time-varying demand.