Large language models are the route to 5G network slice optimization

Commercial context

CSPs must use spectrum as efficiently as possible, yet many devices connected to 5G networks either underuse or overuse available bandwidth. User equipment (UE) devices such as smartphones, robots, and smartwatches for instance often underutilize spectrum. Suboptimal use of networks like this occurs because CSPs lack the key performance indicators needed to allocate different frequency bands effectively across various devices.

The solution

The ‘Unleash the potential of GenAI-powered 5G network slicing’ Catalyst is therefore developing a solution. It involves training large language models (LLMs) to evaluate network performance using 5G data from all spectrum bands. As a result, the model can assess how well network slices meet service level agreements (SLAs) across multiple service categories and usage scenarios. Additionally, it can track network latency experienced by individual devices in real time.

By tackling the challenges of slice management and making full use of all spectrum bands, this solution opens the door to better 5G network monetization. It will also significantly improve customer satisfaction by enhancing performance and reliability. The project is moreover tackling several key challenges in dynamic spectrum sharing (DSS) for 5G network slicing. These include dynamic spectrum allocation, slice orchestration, SLA assurance and automation, public safety, security and AI trust, as well as unclear monetization and adoption strategies.

To address these, the Catalyst begins by training LLMs to evaluate network performance and improve resource allocation. In addition, it supplies AI/ML algorithms with 5G network data across all spectrum bands - including eMBB, URLLC, and mmWave - so they can assess SLA compliance in various scenarios. Moreover, it leverages existing network slicing APIs for service and product orchestration, while also feeding data back to enhance those APIs over time. The solution’s architecture also operates with cloud native solutions using micro service architecture (MSA).

The architecture developed will enable AI algorithms to dynamically share spectrum to ensure each device has the right amount of bandwidth. Ultimately, this approach not only boosts network efficiency but also improves overall user experience.

TM Forum assets used in the solution include TMF 623 SLA Management API, TMF 628 Performance Management API, TMF 656 Service Problem Management, and TMF 675 Geo-Location Management.

Wider application and value

Guru Prasad Uma Maheshwaran, Enterprise Architect at Verizon Communications, one of the Catalyst Champions, notes that the solution improves spectrum efficiency, customer experience, revenue generation, and both energy and cost efficiency. There are more benefits alongside these however. Most importantly, the solution will help CSPs better monetize 5G by more efficiently utilizing all the available spectrum bands to support network slicing.

There are various business and operational benefits on offer here. Firstly, the solution significantly improves spectrum efficiency. It increases spectrum utilization to 15–25% and reclaims 5–10MHz of idle spectrum per 100MHz band. It also boosts QoE and SLA compliance. The system reduces SLA violations by 40–70% and cuts URLLC latency by up to 30%. It reduces manual interventions by 60–80% and shortens the time to resolve RAN bottlenecks from hours to minutes - about 90% faster. It further enhances network throughput – average cell throughput rises by 10–20%, and it offloads 20–40% of 5G NR traffic from LTE.

The solution also delivers strong energy and cost savings. It lowers energy consumption by 10–15% and saves $1M–$5M per region each year on spectrum licensing. Furthermore, it improves customer experience. It reduces churn by 10–25% and lifts Net Promoter Scores by 5 to 15 points. It enables new revenue opportunities as well. SLA-based services drive a 10–20% increase in average revenue per B2B user. The solution also cuts time to market for new slice-based offerings by 50–70%.

Overall, this Catalyst offers considerably improved business outcomes. CSPs can enjoy a 10X faster time to revenue, a 95% faster service activation rate, and a 90% drop in order-to-service OPEX. Besides the direct business benefits, the solution also supports future network optimization efforts. Researchers in 5G, AI, and network slicing can use the generated datasets for analysis, model training, and performance evaluation. These datasets are especially useful for studying resource allocation algorithms. Likewise, network architects and engineers can apply the data to explore diverse scenarios. This will help them design and optimize slicing strategies for a broad range of different applications.