3GPP Release 19 and beyond — AI at the physical layer

Embedding AI in the PHY layer gives the telco ecosystem new tools for efficiency, spectrum use, and performance — and lays the groundwork for 6G’s AI-native vision

5G promised speed and capacity, but delivering on that promise has created new challenges: dense deployments, energy demands, and the limits of traditional optimization methods. 3GPP Release 19 seeks to address these pain points head-on by embedding AI and ML into the RAN and physical layer (PHY), offering operators new tools to tame complexity and unlock efficiency.

AI/ML exploration began in earlier releases, targeting beam management, positioning, and channel state information (CSI) enhancements. Release 19 builds on this work by introducing a general framework for AI/ML on the air interface. Ericsson calls this framework “the backbone for AI integration” into mobile networks. Its potential becomes clear in the concrete use cases Release 19 prioritizes, including:

Beam management

One of the most immediate Release 19 applications of AI/ML is in beam management (BM), which is a technique that enables a base station to efficiently and accurately direct multiple transmission paths to preserve link quality. To help a base station select the optimal beam for a device, the device typically measures the strength of all available beams. When AI is applied at either the device or the network, the best beam can be predicted using data from only a subset of beams, shortening measurement time and reducing unnecessary transmissions. This process improves energy efficiency for both the base station and the device while also conserving valuable radio resources.

As one IEEE research paper noted: “Traditional BM techniques may struggle to keep up with the increasing complexity and dynamic nature of modern communication systems. AI offers the ability to handle this complexity by learning from data and adapting to changing conditions in real-time. This can lead to improved performance, increased efficiency, and reduced costs. Additionally, AI can help in optimizing network resources, enhancing the user experience, and enabling new capabilities that may not be possible with traditional methods.”

Massive MIMO optimization

Massive MIMO promises dramatic gains in capacity and coverage, but managing hundreds of antenna elements is computationally complex. Release 19 introduces AI-driven models that can process high-dimensional channel data more efficiently, optimizing precoding, user pairing, and beam selection in real time. This reduces computational load at the base station while maintaining or even improving performance. For operators, it means extracting the full value of their MIMO deployments, especially in dense environments where multi-user MIMO is key.

Link adaptation and scheduling

Link adaptation and scheduling decisions hinge on choosing the right modulation and coding scheme (MCS) for each user at any given moment. Conventional algorithms rely on predefined thresholds that can be conservative and rigid. By applying AI/ML, Release 19 aims to establish smarter, more context-aware scheduling that adapts dynamically to traffic conditions, user mobility, and interference. The promised result is higher spectral efficiency, more consistent user experience, and improved throughput across the cell.

Signal detection and interference management

As networks densify, interference is increasingly the limiting performance factor. Neural receivers and ML-based equalizers in Release 19 bring data-driven intelligence to the detection process, learning to distinguish desired signals from noise and interference more effectively than traditional approaches. These adaptive receivers can improve reliability in dense urban deployments, extend coverage in weak-signal areas, and reduce error rates, ultimately lowering retransmissions and boosting overall system capacity.

Channel State Information feedback

Underpinning the RAN functions described above, CSI feedback is Channel State Information (CSI) feedback, a process in which user equipment (UE) sends a description of the wireless channel’s current properties back to the base station. The transmitter uses this information to adapt its transmission strategy, adjusting data rate, direction, and signal processing to optimize communication and ensure reliable, high-speed data transfer. 

The traditional process, where UEs measure the downlink channel and feed detailed information back to the gNB, is resource-heavy and slow to adapt. Release 19 introduces AI/ML models that can compress CSI and even predict channel states based on limited measurements. This reduces feedback overhead, shortens response times, and conserves spectrum, while still giving the base station the accuracy it needs for optimal decision-making.

Widespread industry implications

Embedding AI and ML directly into the physical (PHY) layer of the 5G — and eventually 6G — radio stack carries significant implications across the telecom ecosystem.

For vendors, it means designing network equipment that can support not only traditional signal processing but also AI-driven functions for beam management, interference mitigation, and channel estimation. This requires new architectures capable of balancing deterministic PHY operations with adaptive, data-driven models — essentially blending classical engineering with machine learning.

For operators, embedding AI at the PHY translates to greater efficiency and agility in how networks are run. Smarter beam management can extend the reach of mmWave, AI-enhanced link adaptation can improve performance in challenging environments, and reduced measurement overhead can lower power consumption. Together, these improvements can drive down OPEX, help operators make better use of spectrum, and ultimately support new service opportunities.

For chipmakers, the challenge is to integrate AI acceleration at the silicon level without blowing out power budgets. Modems, baseband processors, and RF front-ends must be reimagined to run lightweight inference models at the edge. Success here would give device makers and infrastructure vendors a crucial advantage, while also pushing AI deeper into mass-market hardware.

Release 19 as a bridge to AI-native networks

6G is often described as AI-native — the vision being that AI fundamentally touches every aspect of the network, from planning to optimization. But Release 19, or 5G-Advanced (5G-A), is laying the necessary foundation, bridging today’s practical enhancements and tomorrow’s architectural vision. By embedding AI and ML into the PHY layer and formalizing a framework for lifecycle management, training, and data collection, Release 19 seeks to position AI as more than an experimental add-on and instead, a standardized capability.