#AI-RAN #AI-for-RAN #AI
Apr 28, 2025
SoftBank Corp.
Blogs
AI-RAN aims to build next-generation social infrastructure through the fusion of AI and the Radio Access Network (RAN). SoftBank is working to realize a new form of social infrastructure where AI and telecommunications are seamlessly integrated, through the development of this new architecture. Within this effort, SoftBank defines the initiative to optimize wireless communications using AI as "AI-for-RAN," and is continuously conducting research and demonstrations toward practical implementation.
Within this effort, SoftBank defines the initiative to optimize wireless communications using AI as "AI for RAN," and is continuously conducting research and demonstrations toward practical implementation.
Initial results regarding AI for RAN were presented in a simulation-based format at MWC Barcelona 2024, held in February 2024. Following this, as a further update, SoftBank conducted evaluations in actual over-the-air (OTA) environments and presented significant improvements in communication performance at MWC Barcelona 2025, held in March 2025.
Click below for press release：https://www.softbank.jp/en/corp/news/press/sbkk/2025/20250303_06/
In this article, we will introduce three core technologies identified through the latest demonstration of AI for RAN, detailing their technical backgrounds and potential for social implementation:
1. Uplink Channel Interpolation
2. AI-Driven MAC Scheduling (MAC Scheduler)
3. Sounding Reference Signal Prediction (SRS Prediction)
In 5G networks, communication quality tends to degrade significantly in areas known as "cell edges", which are located far from base stations. This is mainly because uplink signals — transmissions from terminals to base stations — are more vulnerable to effects of distance and obstacles, making accurate channel estimation particularly challenging.
To address this issue, SoftBank developed “UL Channel Interpolation”, a technology that utilizes AI to enhance uplink communication. While last year’s verification was based on simulations, this time, the AI model was operated on a live server and over-the-air (OTA) testing was conducted. Specifically, AI models were embedded into base stations to improve the accuracy of channel estimation during uplink transmissions, enabling precise grasp of complex radio propagation conditions and significantly enhancing communication accuracy.
Figure 1 compares the AI model and the conventional non-AI model in an OTA environment.
In the right-hand side of the figure, the periods with low throughput (indicated by red circles) correspond to cell edge environments, while the periods with high throughput (indicated by blue circles) correspond to cell center environments.
As a result, it was confirmed that, in the OTA testing, the AI model (green line) achieved up to a 50% improvement in communication throughput compared to the non-AI model (yellow line) under cell edge conditions.
Figure 1. Difference in uplink process between AI model and non-AI model
In densely populated areas such as urban centers, where many terminals connect to a base station simultaneously, communication delays and throughput degradation are likely to occur.To address this issue, SoftBank is developing a MAC scheduling technology that leverages MU-MIMO (Multi-User Multiple-Input Multiple-Output). MU-MIMO is a technology that significantly enhances communication efficiency by enabling a single base station to transmit different radio beams to multiple users simultaneously.
SoftBank is currently conducting verification tests to explore whether AI can be used to optimize the dynamic allocation of communication resources—time, frequency, and spatial dimensions—in a 16-layer MU-MIMO configuration.
Figure 2. Overview diagram of MU-MIMO (Multi-User MIMO)
Figure 3 illustrates the concept of MU-MIMO and the operation of scheduling. It shows that multiple users are simultaneously assigned to a single wireless resource composed of time and frequency elements, with AI determining which users to multiplex together.Furthermore, Figure 4 summarizes the throughput improvement in a graph, quantitatively demonstrating about a 10% performance gain. As a result, even during peak congestion periods, individual user wait times are reduced, and the perceived speed of real-time applications such as streaming and video conferencing is improved.
Moreover, by expanding communication capacity, the processing efficiency per base station increases, resulting in a secondary effect of suppressing infrastructure investment.
Figure 3. Scheduling of resource blocks on time (horizontal axis) and frequency (vertical axis)
Figure 4. Results of the effect achieved using MU-MIMO
For users in fast-moving environments—such as in cars or trains—, maintaining a stable and uninterrupted connection is a critical challenge. At the cell edge, communication speed is enhanced through beamforming technology, which focuses radio signals from antennas toward specific terminals by adjusting their directivity. To improve the accuracy of beamforming for terminals that are both at the cell edge and moving at high speeds, SoftBank conducted verification tests utilizing AI to predict the timing when Sounding Reference Signals (SRS) are not transmitted, thereby improving communication performance.
SRS is used by base stations to understand the radio condition between terminals and base stations. However, as the number of users increases, the interval between SRS transmissions becomes longer, making it more difficult to grasp real-time radio conditions. Our AI-based SRS prediction technology utilizes past transmission histories and movement patterns to accurately predict future positions of terminals, enabling optimal beamforming without having to rely on delayed SRS information.Figure 5 presents a dashboard-style comparison of beam control with and without SRS prediction, clearly . showing improvements in communication speed while moving at 80km/h.
This advancement enables stable network connections even for mobile services, and is particularly expected to contribute to fields such as autonomous driving and smart mobility, which are anticipated to see widespread adoption in the near future.
Figure 5. Evaluation results of SRS prediction
These three AI for RAN technologies are all software-driven innovations that can be implemented within the framework of 5G,yet they deliver fundamental improvements in communication quality. Because they do not require any changes on the terminal side, they offer a low barrier to adoption and significant effectiveness, enabling rapid deployment in commercial networks.
Since the announcement at MWC Barcelona 2024, SoftBank has continued to advance evaluations in OTA environments, while also promoting standardization efforts in collaboration with domestic and international telecom operators.Additionally, applications toward network architecture design for the 6G era are being actively explored. AI-based radio optimization is expected to become the "new normal" in the telecommunications industry, and its future developments are drawing increasing attention.
Moreover, the role of AI for RAN extends beyond improving communication quality. By enhancing network capacity through MU-MIMO scheduling optimization, more users can be supported using existing base station infrastructure,which directly helps reduce capital expenditures for telecom operators—a highly significant benefit.
As use cases such as smart cities, autonomous vehicles, drone logistics, and metaverse communication continue to diversify and advance, AI-based control capable of real-time network optimization will become essential. AI for RAN is expected to serve as a foundational technology that addresses this challenge and will play a central role in building next-generation networks.
AI for RAN technologies are expected to play an important role in 6G. SoftBank aims to strengthen its leadership in this area, with a focus on social implementation.