AI-Enabled KPM Granularity Control xApp for O-RAN Digital Twins

6G networks promise ultra-low latency and high throughput which rely on service-layer applications that utilize Artificial Intelligence (AI) and machine learning (ML) for predictive what-if analysis and autonomous decision-making. In next-generation wireless systems, Digital Twins (DTs) have emerged as an enabling technology for this by providing a synchronized virtual replica of the physical network. However, obtaining timely and accurate network data at scale for DTs is challenging due to the massive number of devices and stringent latency constraints. This necessity creates an inherent trade-off in DT communication between synchronization accuracy and the consumption of computation, communication, and storage resources. Furthermore, existing sampling approaches for User Equipment (UE) DTs have not considered the unique constraints of the Open Radio Access Network (O-RAN) standardization efforts where telemetry is generated per-UE based on a pre-defined granularity period. Therefore, in this paper, we propose an xApp that leverages Reinforcement Learning (RL) to dynamically control the per-UE measurement granularity period. This is achieved by interacting with the RAN through standardized E2 service models, specifically E2SM-KPM and E2SM-RC. The resulting three-layered architecture incorporates the O-RAN nodes at the Physical Layer, a distributed message ingestion, processing, and storage system at the DT Layer, and use-case applications at the Service Layer. Experimental results confirm that the proposed methodology balances the trade-off between the fidelity of the twin’s virtual replica and the communication cost required to gather telemetry from the physical network when compared to fixed and heuristic data collection strategies.