Communications

6G AI/ML for Physical-layer: Part I - General Views

By Ameha Tsegaye Abebe Samsung Research
By Feifei Sun Samsung R&D Institute China-Beijing
By Hyoungju Ji Samsung Research
By Eko Onggosanusi Samsung Research America
By Qi Xiong Samsung R&D Institute China-Beijing
By Jaewon Lee Samsung Research
By Wonjun Kim Samsung Research

Introduction

Figure 1. AI enabled modem

For decades, the wireless communication problems have been modeled by statistical models. Consequently, the physical layer of wireless communication systems has relied on linear signal processing techniques. Channel estimation methods such as Least Squares (LS) and Linear Minimum Mean Square Error (LMMSE), linear equalizers, and precoding techniques like Zero-Forcing (ZF) have provided a reliable foundation for 3G, 4G, and early 5G systems. Such linear techniques result in low complexity and near-optimum solutions.

However, the upcoming 6G landscape introduces scenarios that challenge the very assumptions underlying such statistical models:

    
·
Non-linear and time-varying channel environments: With the introduction of a wide range of frequency bands—from sub-1 GHz to upper mid-band and mmWave—the scattering characteristics, delay spreads, and angular spreads vary dramatically. Trying to apply a single linear channel estimation model across diverse frequency bands, user types, and propagation environments leads to performance loss.
    
·
Massive MIMO and dense networks: The dimensionality of channel matrices and interference patterns in massive MIMO systems outgrows the capability of conventional linear estimation and feedback mechanisms. Moreover, beamforming becomes increasingly challenging as the number of users and antennas grows, thereby necessitating adaptive and scalable algorithms.
    
·
Feedback and training overhead: Acquiring full channel state information (CSI) across many antennas and frequency bands requires excessive training reference signals and uplink feedback, thereby limiting efficiency. In FDD systems, the CSI acquisition burden increases linearly with the number of antennas, while in TDD systems, reciprocity calibration may not hold under hardware impairments or mobility – not to mention that sounding reference signal (SRS) is coverage- and interference-limited.
    
·
Lack of contextual adaptation: Linear PHY algorithms operate under fixed rules and cannot dynamically adjust to temporal, spatial, or user-specific conditions. They are unable to leverage past experience, statistical priors, or correlations across time-frequency grids. This limits their ability to generalize or predict.


Why data-driven learning in wireless communication?

The above limitations demonstrate the need for a more flexible, learning-based physical layer. Here, artificial intelligence (AI) and machine learning (ML) solutions are attractive as they do not require a priori statistical models. Instead, AI-based solutions learn the intrinsic features and specific nuances of contexts directly from the data. In this regard, wireless communication systems offer a unique opportunity for the proliferation of AI/ML solutions, as both the network and UEs naturally collect a large volume of training data through measurements. Traditional solutions may still serve as baselines, but they must now coexist or, gradually, be replaced by self-intelligent alternatives that are capable of real-time adaptation.

Table 1. The potential role of AI/ML in improving the wireless communication systems

Table 1 summarizes the potential roles of AI/ML in improving the wireless communication systems. In wireless communication problems that are well characterized by generic statistical models, AI/ML solutions provide improved performance or reduced complexity through context-aware adaptation and joint optimization. When statistical models are unavailable or insufficiently accurate, AI/ML approaches enable practical sub-optimal solutions through data-driven learning. For example, UE power amplifier non-linearities are highly implementation-specific hindering unified statistical modeling.

From 5G to Now

Within 3GPP, the application of AI/ML in the physical layer is a relatively recent endeavor. While early 5G releases explored AI/ML in network management and RAN optimization, until recently, the PHY layer remained untouched due to its stringent latency, complexity, and interpretability requirements.

Starting in Release 18, this began to change. Based on study, three different model usages are discussed. As AI/ML becomes an integral part of the physical layer, it's important to understand the different ways AI/ML models can be deployed between the network and the user device. There are three primary approaches: two-sided, network-sided, and UE-sided models. Each offers a unique trade-off between performance, complexity, and implementation flexibility.

In a two-sided AI/ML model, the data-driven solution is split across the network and the UE for joint optimization. This end-to-end approach allows the network and the UE to collaboratively perform tasks such as channel feedback encoding and decoding. On the other hand, network-sided AI/ML model places the data-driven solution at the base station. The UE behaves as a traditional device, without any AI/ML inference or model awareness. In a UE-sided AI/ML model, the AI/ML model is embedded in the UE. The UE can make self-intelligent inferences locally. The network- and UE-sided models are two sub-categories of the one-sided model.

Figure 2. Operation scenarios for AI/ML in physical-layer

The RAN1 group has initiated studies and undertaken specification work on AI/ML models for physical-layer functions such as:

    
·
CSI feedback compression: CSI compression uses two-sided models. At the UE, it utilizes an auto-encoder to replace the fixed CSI feedback codebooks by adaptive models. The UE encodes high-dimensional CSI into a compressed representation, which is fed back to the network. The network then decodes it to reconstruct the CSI via an auto-decoder. This method can significantly reduce feedback overhead and increase accuracy by learning spatial, frequency and temporal characteristics of the channel.
    
·
CSI prediction: CSI prediction aims to estimate future CSI for the scenarios where CSI feedback is delayed or infrequent. Techniques such as RNNs, LSTMs, and Transformers predict upcoming channel states based on past CSI, enabling more accurate precoding in dynamic environments. Prediction is typically performed at the UE, where channel estimation for the past CSI is available.
    
·
AI-based beam prediction: AI-based beam prediction uses machine learning to predict the optimal beam direction based on sub-sampled measurements. The AI models are trained for optimal beam prediction which implicitly learns mobility patterns, user location and orientation as well propagation environment. It improves beamforming accuracy, reduces measurement overhead, and supports fast beam switching, especially in high-mobility or mmWave scenarios where beam alignment is critical for maintaining link quality.
    
·
AI for positioning: AI/ML-powered localization benefit from context-awareness as the traditional methods are challenged by inaccuracies in time and angle measurements. This involves both "direct positioning" where AI/ML models directly output UE position from channel measurements and "AI-assisted positioning" where AI/ML refines the intermediate measurements used in traditional positioning methods. The key aspects include defining AI/ML model inputs, managing the model life cycle, and supporting various use cases to improve overall positioning reliability and precision.


6G will be the first standard for real AI-based Physical-layer

In 6G, AI/ML is not merely an auxiliary optimization tool—it is being positioned as a core enabler of the radio interface itself. The evolving framework reflects a vision where data-driven intelligence is deeply integrated into the fabric of the wireless protocol stack, spanning both the network and the device. Below is a synthesis of the envisioned direction and structural elements for 6G AI/ML.

Figure 3. Vision of AI/ML in 6G

AI-Native RAN Architecture

The concept of an AI-native self-optimized RAN is central to the 6G vision, leveraging neural networks, federated learning, and predictive analytics to transcend traditional network paradigms. This entails AI not merely assisting radio resource management (RRM) or scheduling but actively replacing or enhancing core PHY/MAC functions—such as MIMO operations optimization, beamforming control, and adaptive link adaptation—through real-time inference and data-driven decision-making. The system must achieve autonomous self-learning, dynamic self-configuration (e.g., via SON and zero-touch network management), and resilient self-healing across protocol layers. This integration positions AI as a foundational component, enabling end-to-end network orchestration and overcoming conventional air interface limitations through AI/ML-driven innovations like adaptive modulation, non-linear signal processing, and cognitive radio techniques.

Framework for 6G AI/ML

6G advances AI integration in wireless networks by refining deployment paradigms introduced in 5G. Two primary classes, one-sided and two-sided models will be the baseline. Rather than rigid classifications, a spectrum of architectures is emerging that leverage model transfer techniques for adaptable and inter-operable AI solutions. These approaches build upon distributed processing capabilities that facilitate the seamless exchange and adaptation of models between network infrastructure (including edge locations) and user equipment (UE).

Far more evolved than the 5G NR AI/ML framework, 6G is envisioned to support advanced AI/ML framework. For example, model transfer with specified model structure may enable site/cell-specific models at the UE. Having been optimized to specific features of a certain cell/site, such localized models may provide better performance or lower complexity than generic models. In particular, a UE-side/part of a model can be partitioned into a fixed part, e.g., fully specified reference model, and an adaptable part (layers). The network may then transfer model parameters for adaptation layers. The approach of transferring model parameters for a limited set of adaptation layers, rather than the full model, effectively mitigates high radio overhead and runtime errors that would otherwise necessitates offline optimization by the UE-side vendor. It is to be noted that transfer of cell/site-specific basis or projection matrices can be considered as linear adaptation which is parallel to model parameters transfer, i.e., non-linear counterpart. This framework can be considered for tasks like CSI compression, where explicit feedback is optimized to significantly reduce overhead compared to conventional codebook-based feedback methods.

In the two-sided model introduced in 5G, standardization becomes paramount for practical implementation in 6G. To accommodate diverse hardware capabilities and interoperability without sacrificing performance, standardized reference models are necessary. The standardized reference model provides a common guidance in the input-output mapping. This enables vendors to train inter-operable models which can also be optimized to the vendors’ respective preferences. The vendors may employ collected field data to optimize their respective models in relation to the relevant additional conditions, e.g., deployment scenarios, antenna layout, and/or model implementation aspects, e.g., complexity and model backbone. As an example, lower complexity is often achieved through techniques like knowledge distillation, quantization and pruning. Standardizing such common reference models is also crucial for ensuring seamless interoperability and integration among various vendors with different proprietary components inherent in their implementations. Furthermore, adhering to minimum performance requirements across various UE and network implementations necessitates a baseline level of consistency enforced by standardized models, ultimately enabling widespread deployment and maximizing the benefits of collaborative AI/ML within the 6G ecosystem. Optionally, network vendors may share information on vendor-specific reference models, e.g., encoder parameters/dataset sharing based approaches. In these approaches, the network may have the freedom in initializing the training process without the constraints of predefined mapping imposed by the reference model- however, these approaches require the UE to implement multiple encoders to interoperate with multiple network vendors.

Data Collection with Privacy Constraints

AI models require real-world data, but 6G must respect privacy, legal, and policy frameworks. Thus, only use-case-specific, purpose-limited data should be collected. Sensitive user data must remain protected, possibly through federated learning or edge-based inference. A common data structure is proposed to support AI, sensing, QoE tracking, and traceability within one cohesive MDT framework.

To successfully embed AI into the radio stack, 6G design should pursue such as moving AI from a peripheral tool to a core part of radio interface logic, combining model-based and data-driven paradigms, supporting lightweight, adaptive, and distributed AI models, and enabling hybrid operation: AI-inferred decisions where possible, traditional signaling where needed.

This direction signals a profound shift from deterministic, rule-based system design toward context-based, self-evolving network intelligence.

Figure 4. Design principles for 6G AI/ML

6G use cases for AI/ML

Two-sided model use cases: In 6G, the two-sided AI model refers to a collaborative framework where both the network and the UE operate on a shared AI-trained representation—most notably for tasks like CSI feedback and precoding. This coordinated structure enables highly efficient compression and improved accuracy compared to traditional signaling methods. The following are the key use cases:

    
·
Explicit CSI Feedback: explicit feedback comprises the UE providing compressed, high-resolution CSI for its own desired downlink channel (termed the explicit channel feedback), and optionally for surrounding interference signals (termed the explicit interference feedback). Unlike traditional codebook-based implicit feedback that only captures a coarse representation of the precoder information conditioned on restrictive transmission hypothesis, this approach supports full channel and interference measurement context not only for ultra-high-resolution precoding, but also for more accurate and flexible link adaptation. One key benefit of explicit feedback in the form of channel matrices is that it facilitates network-side CSI processing by better preserving the phase information. The advanced network-side CSI processing includes precoding and user pairing for multi-user transmission, channel prediction across temporal, frequency and spatial domains as well as precoder determination for non-coherent and coherent joint multi-TRP transmissions.
    
·
Joint Source-Channel-Modulation (JSCM) for CSI: This use case fuses CSI quantization, channel coding, and modulation into a jointly optimized AI-based pipeline. The CSI compression is considered as source coding as it aims to obtain a compact representation of CSI. On the other hand, the channel coding provides redundancy to mitigate errors due noise and interference. Instead of handling each function separately, the system learns how to minimize the total end-to-end distortion, thereby achieving better robustness in harsh channel conditions, reducing cumulative quantization and coding losses, and enabling seamless uplink resource allocation with minimal redundancy. This approach is also be referred to as semantic CSI feedback, where the end-to-end optimization focuses on preserving the underlining meaning (semantics) of the channel information, allowing some degree of distortion- unlike conventional channel coding, which aims for perfect reconstruction despite the lossy CSI representation.
    
·
TPMI Compression and Feedback Optimization: In frequency-selective massive MIMO scenarios, overhead for transmit precoding matrix indicators (TPMI) can be large and inefficient to signal. The two-sided model supports both AI-based encoding of TPMI using a compact index space and effective support for sub-band-wise precoding, and better scalability with multi-layer transmission. If alignment between the reconstructed TPMI at the UE auto-decoder and the uncompressed TPMI at the network auto-encoder can be guaranteed, this is a good enhancement for fixed wireless access (FWA) systems where coherent transmission from relatively large number of antenna ports (up to 8 for FR1, or 16 for FR3) is essential.
    
·
PAPR reduction for waveform precoding: To mitigate the residual peak-to-average power ratio (PAPR) challenge in uplink DFT-s-OFDM transmissions, we explore a novel AI-driven approach leveraging a two-sided model architecture. This framework employs deep-learned transform precoding matrices specific to data symbols to minimize PAPR and improve transmission efficiency. The resulting PAPR reduction enhances uplink coverage and overall system performance by enabling the UE to transmit at higher power levels.


NW-side model use cases: these refer to scenarios where the data-driven intelligence is fully located at the network, without requiring any AI model coordination or execution at the UE. These models are particularly attractive because they reduce device complexity and power consumption while still enabling advanced PHY-layer functionalities through intelligent inference at the network side. Here, network-side data collection simplifies site/cell-specific deployment, unlike UE-side models that require frequent model switching due to user mobility. Some of the envisioned use cases for NW-side models are outlined below:

    
·
CSI Feedback Using AI-Trained Basis: With a shared AI-trained basis (e.g., to derive W1 matrix), the UE encodes high-resolution downlink channel state information (CSI) into a compressed form, which the network can decode using the same model. This approach enables ultra-compressed, high-precision CSI reporting.
    
·
SRS channel denoising/reconstruction: In traditional TDD systems, the network relies on uplink SRS (Sounding Reference Signals) to estimate the downlink channel through reciprocity. However, when SRS is unavailable or degraded—such as in low-power, interference-limited, or inactive carriers—an AI model can infer the downlink CSI based solely on partial or historical information at the network This enables feedback-free channel prediction, thereby conserving UE energy and uplink resources. Additionally, the network-side model may correct SRS channel distortion that would result from SRS antenna power imbalance due to insertion loss. In particular, the NW-side model can fuse power-imbalance-free explicit CSI feedback corresponding to a subset of UE’s RX ports with SRS measurement transmitted by the complementary subset of RX ports.
    
·
NW-side Joint Source-Channel-Modulation (JSCM) for CSI: JSCM can also be realized by NW-side model. Particularly, the UE projects the measured channel into a lower-dimensional space using a linear-projection matrix. Accordingly, CSI encoding would constitute solely linear projection. The projection matrix, which is also referred as measurement/sensing matrix in the compressive-sensing lexicon, can be AI-trained, generated from a known-sequence or sampled from random distribution, e.g., Gaussian. The network-side model which is trained with the knowledge of the projection matrix and field dataset would reconstruct CSI from ‘analog feedback’. This approach avoids AI/ML model inference at the UE, thereby reducing UE-side computational complexity compared to two-sided model based JSCM.


UE-sided model use cases: UE-sided AI models refer to cases where intelligence is embedded in the UE, thereby enabling local inference and decision-making based on UE’s conditions. One advantage of UE-side inference is UE’s access to high-resolution measurements from downlink reference resources that are less constrained by coverage limitations and quantization error. Additional key use cases are presented below

    
·
Cross-frequency CSI Prediction in 6G: Cross-Frequency CSI prediction uses AI to estimate unmeasured channel information across sub-bands, bandwidth parts and carriers, reducing the need for full-band CSI-RS. Here, the model learns frequency correlation to interpolate or extrapolate the missing CSI, this enables fast CSI acquisition in yet activated BWP, CCS as well as reduces CSI overhead (especially in wideband and CA), and, thereby energy saving with carrier off.
    
·
CSI-RS overhead reduction and AI/ML-based Channel Reconstruction: The network transmits a spatially sub-sampled CSI-RS CSI rather than full-dimension channel probes. This technique greatly reduces downlink overhead while maintaining performance—especially beneficial in massive MIMO scenarios with hundreds of antenna ports. Using a trained AI model, the UE then reconstructs the complete channel state with high accuracy from sub-sampled measurements, e.g., from 64 ports measurement to 256 ports. The same approach can be applied to frequency and temporal domain too through sub-sampled measurement and reconstruction to ultimately reduce CSI-RS overhead and energy consumption.


AI-receiver: the physical layer must handle increasingly complex and non-linear environments, where traditional receivers such as LMMSE and MMSE-SIC reach their performance limits. The AI receiver emerges as a key innovation to overcome these challenges by leveraging deep learning models to enhance signal detection and channel estimation in difficult scenarios.

Specifically, AI receivers are designed to:

    
·
 Perform robust channel estimation even under sparse DMRS configurations or severe power imbalance between transmission paths. Neural networks can interpolate missing channel information and compensate for pilot degradation.
    
·
 Handle non-linear distortions at the base station (BS) receiver, which become more prominent when user equipment (UE) transmits at high power. AI models can learn and mitigate hardware non-idealities that are difficult to model analytically.
    
·
By integrating AI into the receiver chain, 6G systems can achieve improved decoding accuracy, extended coverage in low-SNR environments, and reduced reliance on dense reference signaling. The AI receiver is a critical enabler for reliable and efficient communication in next-generation networks.


In the sequels of this blog, we will elaborate on the details of use cases we are hoping to use in 6G.

Conclusion

The paradigm motivated by statistical modeling of the wireless communication system has been a faithful companion in its evolution, but it is approaching its practical limits. As 6G aims to support ultra-dense networks, massive-scale antenna arrays, extreme mobility, and integrated sensing, a new foundation is needed.

AI/ML offers a path forward by enabling a physical layer that is context-aware, predictive, and adaptive through data-driven learning. This transformation is already underway in 3GPP, where physical layer studies are now incorporating deep learning models into critical measurement, feedback and estimation procedures. To ensure that this vision comes true, a paradigm shift is essential – not only in algorithm design, but also in architecture, standardization of model interfaces, development of lightweight inference engines, and perhaps most critically, a rethinking of the PHY layer’s role as not just a signal processor—but an intelligent agent in the network.

The 6G physical layer will no longer serve as a bit delivery engine, but it will also learn, adapt, decide, and, furthermore, “think”. That is the promise of AI-native PHY.