As 5G mobile networks grow in complexity, it’s becoming impossible for human network engineers to manage the network without automation. The evolution from rules-based automation to the use of AI/ML is creating a category of management solutions that can anticipate problems in the network and fix them in real time.
Digital twins are broadly used in complex digital systems that need an exact mirror image of the digital environment in all its complexity so that testing and optimization can be done without disrupting the real-world system at work.
In mobile networks, the use of digital twins provides the modeling and simulation capabilities to train these AI-based network management systems.
Digital twins and RICs working together
To add new management functionality and control to Open RAN, the O-RAN ALLIANCE created standards for the non-real-time (non-RT) and near real-time (near-RT) RAN intelligent controllers (RICs).
RICs take the role of central controllers for network operations, including the tasks related to radio resource management. The operational tasks are optimized by employing the right apps in the RICs, wherein AI/ML becomes a powerful tool that can potentially tackle many complex problems that are difficult or unsolvable today.
For calibration, AI/ML models require plenty of data, which may not be available at the initial phase of network rollout. The lack of data for training purposes can be overcome by a digital twin, whose models can generate data by themselves.
Then, the digital twin evolves with the expansion of the real network and there is a creation of a cycle of data exchange: 1) the digital twin supplies meaningful data to RICs to train AI/ML models, which are used to infer optimal network configurations; 2) RIC feeds the updated network configurations to the real network to maintain network operations; and 3) the data reflecting the health and efficiency of the real network are gathered to recalibrate the models in the digital twin.
There are three building blocks for creating an Open RAN digital twin:
- The modelling entities that create accurate digital replicas of different aspects of a RAN network. Data captured from the Open RAN standardized interfaces (O1, O2, E2 and A1) is used to synchronize the models with the live physical network.
- The RAN Scenario Generator is powered by AI/ML technology and automatically parameterizes the models to generate billions of training scenarios for the AI/ML models. The RAN Scenario Generator can also automatically evolve itself based on the performance feedback from the RAN analytic module, generating more and more challenging training data set for the AI/ML intelligence and performance to continuously improve.
- Advanced Visualization simplifies the data and presents it to network engineers when their input is needed.
Examples of modelling entities are:
- Modeling Physical RF Propagation: Ray tracing is increasingly being used to realistically model the physical RF propagation characteristics in the mobility/RF model used by the network. This technique estimates the RF propagation characteristics and impact of buildings and other obstructions based on calculating the path gains of propagation paths through a geometrical region of varying velocity, absorption characteristics, and multiple reflecting surfaces.
Both the ray tracing algorithm and the measurement data that calibrate the penetration loss, reflection and scattering characteristics of the surfaces have a bearing on the accuracy of the calculations. Using a digital twin provides a large amount of information that is constantly updated enabling these calculations to be made much more accurately.
- Modeling RAN and cloud: Modeling RAN and cloud-based network elements is challenging because various systems and resources have dynamic behaviors. The model has to account for both the changing infrastructure and has to stay synchronized with real-time states of the physical network. This requires a model that relaxes the real-time constraints of the digital twin to near real time which reduces computing complexity. Alternatively, the digital twin can utilize a GPU-based cloud service that provides the additional compute needed for this complexity. Either way, the digital twin enables the network, UE state, call flow and KPI prediction capabilities needed for this modeling.
Three emerging use cases
How can digital twins be used today? Here are some use cases where the interplay between digital twin and AI/ML models in RICs is essential:
Network Energy Saving There are exciting new capabilities to adjust power levels of network elements when traffic is low in order to save energy. All of these methodologies use historical information to balance power consumption with maintaining performance – a good application for a digital twin.
Massive multiple-input multiple-output (mMIMO) antennas and network densification, for example, are implemented to improve performance of ultra-reliable and low latency communications (uRLLC), mobile broadband (MBB), and machine-type communications (MTC). But mMIMO takes more compute power increasing an MNO’s carbon footprint.
To minimize the extra power needed, mMIMO antennas can be downgraded by turning off unneeded RF circuits during low traffic periods. Similarly, certain cells or carriers can be switched off during low traffic hours. Using power management capabilities built into Intel architecture processors can further reduce power consumption by switching off or reducing CPU cycles of other RAN elements such as distributed units (DUs) and centralized units (CUs). With a digital twin providing historical information, the RIC can downgrade performance of the antennas when traffic is low saving energy.
mMIMO Beamforming Optimization Fully digital mMIMO may be too costly for some sites, especially those sites where a high carrier radio frequency is used. The alternative solution is using hybrid antenna arrays, which still can exploit the large degrees of freedom a mMIMO antenna system can offer.
However, without full digital control of each antenna element these antennas are controlled in groups. As such, the mobile terminals in a cell are served by a grid-of-beams – those beams of different directions and widths are semi-static and designed to cover a geometric region of cell.
But mobile terminals are never distributed uniformly in an area. A grid-of-beams can be optimized to provide better experience to the users. The optimization can be done iteratively with the help of AI/ML. Any change of the user distribution will cause a change of grid-of-beams optimization as the AI/ML model is able to represent the outcomes of the optimized configuration.
Zero Touch Network Management The complexity of today’s networks makes human management overwhelming. Zero-touch network management is in high demand as no human error is possible when optimized configuration rules are used. The centralized control offered by RICs, with vital AI/ML input, makes RAN and the whole radio communication system more intelligent, hence, brings in benefits in both economic and environmental aspects.
Conclusion
Network complexity is here to stay, driving MNOs to embrace automation and AI/ML tools to keep the network throughput optimized and to ensure the minimum energy use for cost savings and sustainability. The use of digital twins makes available an important diagnostic tool for Open RAN and other network elements.
With digital twins in the system, MNOs have a new way to predict the behaviors of the real network. They can also use digital twins for low cost maintenance of, and quick response to, faults due to the digital twin’s highly accurate replica of the physical network. As the three use cases in this paper show, digital twin systems add a lot of value to network deployment and operations.
