Skip to main content
Publications lead hero image abstract pattern


Written By:

Xingqin Lin, NVIDIA, USA

Published: 25 Mar 2024


CTN Issue: March 2024

A note from the editor:

For the month of March 2024, CTN will dive into what else?  AI, in this case as it relates to Radio Access Networks (RAN) 3GPP Standardization and what transpired in the recently completed R18.  As the telecommunications landscape continues to evolve at a rapid pace, and as the telecommunication’s industry sets it sights in completing work in 5G Advanced, followed by 6G studies and eventually 6G normative work, we see how the talk of AI being fundamental to the future of wireless communications continues to increase.  The last couple of months we have seen several announcements to that effect: AI for RAN, the telco’s AI Alliance, and NVIDIA’s 6G Research Cloud Platform all have the common theme of exploring how AI can become a fundamental capability in the wireless communication evolution.  But before we get there, AI must prove its metal, and 3GPP is already hard at work investigating possible applications of AI in each release.  In this month’s article the author provides an overall review of what was done in R18 as it relates to a general framework and discusses several key use cases that were part of the release work.  There’s no doubt that the AI work in 3GPP it is still in its infancy, but there’s also no doubt that the expectation for what AI can do in cellular communications is high, only time will tell us how deep and wide it can go.

Miguel Dajer, CTN Editor-in-Chief

An Overview of the 3GPP Study on Artificial Intelligence for 5G New Radio

Xingqin Lin

Xingqin Lin


1. Introduction

As the first release of 5G-Advanced, the 3rd Generation Partnership Project (3GPP) Release 18 includes comprehensive study and work items which not only cater to immediate commercial needs but also encompass long-term endeavors that lay the groundwork for 6G [1][2]. The integration of artificial intelligence (AI)/machine learning (ML) into the fabric of 5G-Advanced standards evolution is set to facilitate the widespread adoption of AI/ML in commercial wireless systems. Prior to the advent of 5G-Advanced evolution, 3GPP engaged in preliminary AI/ML initiatives during the first phase of 5G evolution, spanning multiple domains from the 5G core network, the operations, administration, and maintenance (OAM), and the radio access network (RAN) [3][4].

Air interface is a fundamental component in any wireless communication system. Recent works such as [5] have painted visions of an AI-native air interface. The preliminary AI/ML initiatives conducted by 3GPP prior to the 5G-Advanced evolution did not cover the 5G New Radio (NR) air interface. 3GPP addressed this gap in Release 18 by exploring the potential of using AI/ML-based algorithms to enhance the NR air interface [6]. This is the first of its kind in the 3GPP’s development of wireless communication standards. The scope of the 3GPP study on AI/ML for NR air interface included the development of a general AI/ML framework, alongside the exploration of specific use cases such as channel state information (CSI) feedback, beam management, and positioning [7]. In this article, we provide an overview of the 3GPP Release-18 study on AI/ML for the NR air interface.

2. General AI/ML Framework

To establish a general 3GPP AI/ML framework for the air interface, considerable efforts from 3GPP members have been dedicated to formulating common terminology that pertains to AI/ML functions, procedures, and interfaces. The AI/ML functional framework defined by 3GPP for the NR air interface includes a set of core functions, including data collection, model training, management, inference, and model storage.

An AI/ML model needs to be developed, deployed, and managed during the entire lifecycle—a process known as AI/ML model life cycle management (LCM). 3GPP has studied two distinct methods for managing the life cycle of an AI/ML model at user equipment (UE). The first method is categorized as functionality-based LCM. A functionality refers to an AI/ML-enabled feature or feature group facilitated by a configuration. AI/ML functionality identification fosters mutual understanding between the network and the UE about the AI/ML functionality. The process of functionality identification may be integrated within the existing UE capability signaling framework. Essentially, configurations are tailored in accordance with conditions indicated by UE capability. Subsequently, upon identifying functionalities, the UE can report updates about the applicable functionalities among those configured or identified. In functionality-based LCM, the network indicates selection, activation, deactivation, switching, and fallback of an AI/ML functionality through 3GPP signaling. Notably, the exact AI/ML model(s) that underpin a given functionality might not be identified at the network.

The second method is categorized as model identity (ID) based-LCM. A model ID serves as a distinctive identifier for an AI/ML model, wherein the model could be logical and its mapping to a physical model is up to implementation. AI/ML model identification ensures a mutual understanding between the network and the UE concerning the AI/ML model in question. Specifically, the AI/ML model is identified by its designated model ID at the network, and the UE indicates its supported AI/ML model to the network. Besides the model ID, the model can have accompanying conditions as part of the UE capability definition as well as additional conditions (e.g., scenarios, sites, and datasets) which determine the applicability of the model. In model-ID-based LCM, both the network and the UE may perform selection, activation, deactivation, switching, and fallback of an AI/ML model by using the corresponding model ID.

The commercial deployment of an AI/ML-enabled feature hinges on its ability to deliver reliable performance across different scenarios, configurations, and site-specific conditions in mobile communication systems. To achieve this objective, 3GPP has investigated three approaches: model generalization, model switching, and model update. Model generalization aims to develop one model generalizable to different scenarios, configurations, or sites. Alternatively, a set of specific models can be developed—ranging from scenario-specific to configuration- or site-specific. Within this set of models, the technique of model switching is harnessed to effectively address the different scenarios, configurations, or sites. The process of model update, often involving fine-tuning, entails a flexible adaptation of the model structure or its parameters in response to changes in scenarios, configurations, or sites. A pivotal principle underpinning these approaches is to ensure that the performance of AI/ML-enabled features remains at a level equal to or better than that of legacy non-AI/ML-based operations. Therefore, performance monitoring is a must for the AI/ML-enabled features, calling for functions such as computing monitored performance metrics, reporting monitoring results, and control signaling mechanisms to swiftly recover from failure.

Different use cases require varying degrees of collaboration between the network and the UE for the corresponding AI/ML operations. 3GPP has identified three distinctive levels of network-UE collaboration:

  • Level x–no collaboration: AI/ML operations are up to proprietary implementations, which do not require any specific standards enhancements tailored for AI/ML functionalities.
  • Level y–signaling-based collaboration without model transfer: AI/ML operations integrate dedicated standards enhancements to facilitate the process without involving model transfer. Here, ‘model transfer’ refers to the delivery of an AI/ML model over the air interface from one entity to another, conducted in a manner not transparent to 3GPP signaling mechanisms.
  • Level z–signaling-based collaboration with model transfer: AI/ML operations encompass not only the integration of new signaling but also leverage advanced model transfer capabilities.

In essence, these collaboration levels encompass a spectrum from minimal involvement to deep integration, signifying the versatility and adaptability of AI/ML operations across different contexts.

Figure 1: Types of AI/ML model training for CSI compression using a two-sided model.
Figure 1: Types of AI/ML model training for CSI compression using a two-sided model.

3. Use Case: CSI Feedback

CSI refers to the information of the multipath wireless channel between a 5G node B (gNB) and a UE. The UE can measure downlink reference signals, compute downlink CSI, and provide a CSI report to the gNB, thereby facilitating downlink transmission. 3GPP has studied two representative sub-use cases of employing AI/ML-based algorithms for CSI feedback enhancement: spatial-frequency domain CSI compression and time domain CSI prediction with a UE-sided model.

In AI/ML-based CSI compression, a UE employs an AI/ML-based CSI encoder to generate CSI feedback information, while a corresponding AI/ML-based CSI decoder at the gNB is used to reconstruct the CSI from the received feedback data. This is an example of a two-sided AI/ML model, where the inference operation is split between the UE and the gNB. This two-sided AI/ML model can be employed to compress either the raw channel matrix estimated by the UE or the precoding matrix derived from the raw channel matrix. Notably, compressing the precoding matrix aligns with the existing codebook-based CSI feedback framework specified for the NR air interface, thus attracting more interest during the 3GPP study.

Using a two-sided AI/ML model for the air interface introduces a multitude of challenges. The first challenge involves the training of the two-sided AI/ML model. Within this context, 3GPP has investigated three types of training that involve varying degrees of collaboration between the network and the UE, as illustrated in Fig. 1. The training complexity inherent in a two-sided AI/ML model for the air interface is further compounded by the necessity for multi-vendor interoperability and compatibility. The CSI decoder located at the gNB needs to be compatible with different CSI encoders at the UEs, and vice versa. In scenarios where a common CSI decoder model is utilized for multiple CSI encoder models, the network-side training entity—under training type 1 or 2—must coordinate with UE vendors for joint training efforts. The release of a new UE type could potentially trigger retraining across all vendors. Similar challenges exist in training type 1 or 2 when a shared CSI encoder model is used for multiple CSI decoder models. These issues can be mitigated in training type 3. In particular, if a common CSI decoder model is used for multiple CSI encoder models (or vice versa), the retraining due to the release of a new UE type can be confined to involve only the associated UE vendor and the network vendor due to the separate training nature in training type 3.

A challenge in the legacy CSI reporting framework of NR pertains to a temporal lag between the time to which the reported CSI corresponds and the moment at which the gNB actually employs the CSI report. This time delay leads to a situation where the reported CSI becomes outdated, a phenomenon commonly referred to as channel aging. The pace at which the reported CSI becomes outdated is amplified by higher UE speeds. This concern becomes particularly pronounced in the context of multi-user multiple-input multiple-output (MU-MIMO) in massive MIMO deployments. The performance of MU-MIMO has been observed to deteriorate when UEs move at medium to high speeds.

In contrast to CSI compression, which necessitates a two-sided model, time domain CSI prediction can employ a one-sided model. Training this one-sided model can be executed by a single vendor, and the inference can subsequently be conducted by one side (either the gNB or the UE). Considering the workload in Release 18, 3GPP chose to focus on the UE-sided model for time domain CSI prediction. From the standpoint of 3GPP standards, it is expected that we can largely reuse the existing CSI framework to support CSI prediction. In particular, the AI/ML model LCM for UE-sided CSI prediction can to a large extent reuse what is defined for other UE-sided use cases.

4. Use Case: Beam Management

Beam management functionality in NR is used to support beamforming. It is particularly needed for 5G millimeter wave systems that rely on analog beamforming. In a basic procedure for downlink beam management, the UE measures the reference signal associated with each gNB’s transmit beam and tests different UE receive beams for each gNB’s transmit beam to find a suitable downlink beam pair. This process can be time-consuming and entail a substantial overhead in terms of reference signals.

3GPP has studied two representative sub-use cases that involve the application of AI/ML-based algorithms to beam management. They are referred to as ‘spatial-domain downlink beam prediction’ and ‘temporal downlink beam prediction.’ Spatial-domain downlink beam prediction leverages measurement outcomes from a designated set of downlink beams, denoted as ‘Set B,’ to predict the best beam within another set of downlink beams, termed ‘Set A,’ at the present moment. Temporal downlink beam prediction harnesses historical measurement results derived from ‘Set B’ to anticipate the best beam in ‘Set A’ for one or more future time instances.

A typical input to an AI/ML model for the spatial-domain or temporal downlink beam prediction is layer 1 reference signal received power (L1-RSRP) measurements of beams within ‘Set B.’ A typical output from the AI/ML model is the predicted top-K beams in ‘Set A.’ The AI/ML model training and inference can reside at the gNB side or at the UE side. In scenarios where AI/ML inference occurs at the UE side, the UE needs to report its predicted beam(s) to the gNB. Alternatively, when AI/ML inference takes place at the gNB side, the UE is required to report its L1-RSRP measurements for the beams within ‘Set B’ to the gNB.

To provide clarity regarding AI/ML-based downlink beam prediction, we now elaborate on how beam prediction can be integrated into the existing NR beam management framework. Recall that the existing framework consists of three procedures known as P1 (initial beam pair establishment), P2 (transmit beam refinement), and P3 (receive beam refinement). Figure 2 provides an illustration of downlink beam management procedures with AI/ML-based beam prediction at the gNB side and at the UE side.

Figure 2: Downlink beam management procedures with AI/ML-based beam prediction at gNB side (left) and at UE side (right).
Figure 2: Downlink beam management procedures with AI/ML-based beam prediction at gNB side (left) and at UE side (right).

5. Use Case: Positioning

The existing 5G NR positioning methods are typically geometry-based, consisting of two main steps: 1) conducting measurements of radio signals, and 2) calculating a position estimate by solving a system of non-linear equations that establish a relationship between the UE’s position and the measurements. The accuracy of the geometry-based positioning methods heavily depends on the availability of measurements linked with line-of-sight (LOS) paths. In scenarios involving weak LOS conditions or dense multipath environments, such as indoor factory settings, the accuracy of geometry-based methods tends to degrade.

3GPP has studied two representative sub-use cases involving the application of AI/ML-based algorithms to positioning. These have been termed as ‘direct AI/ML positioning’ and ‘AI/ML-assisted positioning.’ Direct AI/ML positioning employs an AI/ML model to directly determine the location of UE. For instance, this can encompass fingerprinting-based positioning utilizing channel observations, such as channel impulse response or power delay profile, as input to the AI/ML model. AI/ML-assisted positioning involves leveraging an AI/ML model to generate an intermediate measurement statistic, which is instrumental in positioning. This could encompass measurements such as LOS probability, angle-of-arrival/departure, or time-of-arrival.

The AI/ML model training and inference can reside at the UE side, the location management function (LMF) side, or the gNB side. Depending on the roles of UE, LMF, and gNB in the positioning procedures, 3GPP focused on three categories of positioning methods. The first category is UE-based positioning, where the UE itself executes either direct AI/ML positioning or AI/ML-assisted positioning. The second category is UE-assisted LMF-based positioning, where the UE provides assistance to the LMF in estimating the UE’s location. In this scenario, the LMF can perform direct AI/ML positioning, or the UE can engage in AI/ML-assisted positioning. The third category is gNB-assisted positioning, where the gNB provides assistance to the LMF in estimating the UE’s location. In this case, the LMF can implement direct AI/ML positioning, or the gNB can participate in AI/ML-assisted positioning.

Figure 3: Reference block diagram for testing AI/ML-based features.
Figure 3: Reference block diagram for testing AI/ML-based features.

6. Interoperability and Testability

Interoperability and testability are critical considerations in the development and deployment of standardized features within cellular networks, including AI/ML-based schemes. 3GPP RAN working group 4 (RAN4), responsible for setting performance requirements and defining test procedures, is investigating the interoperability and testability aspects for validating AI/ML-based performance enhancements. The incorporation of AI/ML into the air interface introduces significant challenges to the existing requirements and testing framework. AI/ML models are data-driven and often lack physical interpretations, rendering the prediction of their performance difficult. In this section, we highlight the key areas under development in 3GPP.

The scope of 3GPP RAN4 requirements and testing for AI/ML-based features encompasses a range of vital elements, including inference, LCM procedures, data generation and collection, and generalization verification. Core requirements include the performance monitoring procedure, functionality/model management procedure, and the corresponding latency and interruption requirements. 3GPP considers a reference block diagram for testing AI/ML-based features, as illustrated in Fig. 3. Within this framework, the device under test (DUT) can be either UE or gNB. The reference block diagram covers both one-sided and two-sided models. In the latter case, the test equipment incorporates a companion AI/ML model to perform joint inference with the model at the DUT. However, the methodology for devising a reference AI/ML model within the testing equipment to effectively test the performance of the corresponding AI/ML model within the DUT remains an ongoing topic of discussion.

While the two-sided models pose more interoperability challenges, it is important to note that interoperability considerations also extend to the utilization of one-sided AI/ML models. This includes aspects such as procedure signaling and testing setups to ensure compliance with minimum requirements. Traditionally, 3GPP RAN4 defines requirements for testing equipment in controlled laboratory conditions prior to its field deployment. Generalization verification is an intricate task. In particular, the training of an AI/ML model can be tailored to encompass the entirety of conditions outlined in the standard, thus demonstrating superior performance during testing. However, the AI/ML model may be overfitted to the standardized setups, possibly compromising its robustness in real-world environments that are not aligned with the controlled conditions. Moreover, AI/ML models might need periodic updates even after they are deployed in live networks. These new considerations call for the implementation of performance monitoring mechanisms to detect non-compliance, as well as the formulation of new testing procedures to effectively validate the functionality of AI/ML-based features operating in the field.

7. Conclusion and Future Outlook

The 3GPP Release-18 study on AI/ML for the NR air interface is a pioneering initiative in the 3GPP’s development of wireless communication standards. This article has offered a timely overview of the key topics investigated by 3GPP in this area. Considering that this is a largely uncharted territory for standards development, substantial work remains ahead within 3GPP to cultivate this domain into a state of maturity fit for commercial deployments at scale. In the upcoming Release 19, 3GPP will conduct normative work to introduce support for the AI/ML general framework for one-sided AI/ML models and specify necessary signaling and mechanisms to enable AI/ML-based beam management and positioning [9]. Additionally, new use cases such as AI/ML-based mobility management will be explored, and further studies will delve into areas that warrant deeper investigation, such as CSI feedback and testing methodologies for two-sided AI/ML models. These concerted efforts are poised to lay the foundation for the forthcoming 6G that will feature integrated AI and communication as a key usage scenario.


  1. X. Lin, “An overview of 5G Advanced evolution in 3GPP release 18,” IEEE Communications Standards Magazine, vol. 6, no. 3, pp. 77-83, Sep. 2022.
  2. W. Chen et al., “5G-Advanced towards 6G: Past, present, and future,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 6, pp. 1592-1619, Jun. 2023.
  3. X. Lin, “Artificial intelligence in 3GPP 5G-Advanced: A survey,” IEEE ComSoc Technology News, Aug. 2023.
  4. 3GPP TR 37.817, “Study on enhancement for data collection for NR and EN-DC,” V17.0.0, Apr. 2022.
  5. J. Hoydis et al., “Toward a 6G AI-native air interface,” IEEE Communications Magazine, vol. 59, no. 5, pp. 76-81, May 2021.
  6. 3GPP TR 38.843, “Study on artificial intelligence (AI)/machine learning (ML) for NR air interface,” V0.1.0., June 2023.
  7. RP-213599, “Study on artificial intelligence (AI)/machine learning (ML) for NR air interface,” 3GPP TSG RAN Meeting #94e, Dec. 2021.
  8. X. Lin et al., “Embracing AI in 5G-Advanced towards 6G: A joint 3GPP and O-RAN perspective,” IEEE Communications Standards Magazine, vol. 7, no. 4, pp. 76-83, Dec. 2023.
  9. X. Lin, “The bridge toward 6G: 5G-Advanced evolution in 3GPP Release 19,” arXiv preprint arXiv:2312.15174, Dec. 2023.

Statements and opinions given in a work published by the IEEE or the IEEE Communications Society are the expressions of the author(s). Responsibility for the content of published articles rests upon the authors(s), not IEEE nor the IEEE Communications Society.

Sign In to Comment