This year, we expect to see significant progress in one of the technology shifts required to implement 6G: the rise of AI-native RANs.
Currently, AI is used as an add-on, an optimization tool that – while beneficial – is not central to the network. The 6G network’s AI-native architecture will instead use it like the central nervous system of the network to give intelligence for every decision. This transition enables autonomous operations across complex, heterogeneous environments managed by programmable platforms like the RAN Intelligent Controller (RIC).
However, this shift is not without challenges, with AI models being prone to drift and becoming unreliable when scaling across dynamic, real-world conditions. This degradation happens because models reflect the world as it existed during training, while real-world threats and conditions continue to evolve. Even in an ideal scenario, where training data provides a complete picture of the entire network rather than just fragmented snapshots from siloed departments, that data remains inherently historical. In practice, even in large organizations with good communication flows, training data can be just a glimpse in the rearview mirror.
RAN digital twins enable a hybrid data strategy, where AI algorithms can be continuously recalibrated and tested using a mix of both emulated traffic and real-world data.
Through this, it’s possible to train for potential upcoming threats, such as DoS attacks, as well as possible changes before they happen in the real network.
Real versus synthetic training data
The first step in creating the twin and AI models running in it is training. Any model is only as good as the data on which it is trained. Access to reliable, high-quality data is therefore critical to ensure the twin mimics the real-world network as closely as can be achieved.
The traditional approach is to use data from the live network itself: performance management counters, call traces, and logs, all extracted from network elements and operational support systems. But this comes with several notable limitations.
It is historical only, and therefore, new attacks or unencountered edge cases cannot be planned for. It is limited: at best, it is a snapshot of network traffic across one network. In reality, it is more likely to be a subset of a subset of the snapshot, with data provided by just a few divisions within the operator. For third-party developers of Open RAN equipment, live network data is often impossible to access.
To overcome these weaknesses, the alternative approach is to use synthetic data, created using an AI RAN Scenario Generator (RSG). This emulated data approach has no delays in access and no security, privacy, or commercial sensitivities. It can predict interference from the building of new towers, new buildings, and the use of new wireless frequencies – e.g.6G’s FR3 band. It can build in as-yet-unseen DoS and DDoS attacks, as well as edge cases, and extreme traffic levels. And it can also be integrated with key real-world data vectors to give a true representation of an individual network to create the ideal training platform: a hybrid data layer.
Figure 1. The AI RAN scenario generator uses the hybrid data layer to train and challenge xApps and rApps for use in digital twins. (Image: VIAVI)
The hybrid data layer
Through this method, it is possible to achieve a high-fidelity digital twin that mirrors the physics and behavior of the specific real-world network.
This twin environment can then be used to generate data that is not only realistic but also forward-looking. This enables training of agentic AI-based workflows and AI models of the x (near-real-time) and r (non-real-time) apps used by the RIC to run the RSG-created hypothetical scenarios.
At this point, it should also be noted that it is not simply enough to emulate user equipment and traffic profiles. The location in which it is being used is also critical. A hybrid data layer approach should therefore also look to ray-tracing tools to incorporate the position of towers within a specific area and to capture signal propagation patterns around hills, trees, buildings, and traffic profiles — all enhancing real-world accuracy.
Figure 2. A digital twin with ray tracing captures, shown by VIAVI at MWC 2026 in partnership with NVIDIA as part of its Aerial Omniverse Digital Twin platform. (Image: VIAVI)
Combating AI drift
Networks are far from static entities, and any AI model and digital twin that are left alone will quickly become out of sync with the real world. This AI drift can lead a model to make poor decisions with unintended consequences. And drift will be particularly pronounced if the end goal is poorly defined, with the application achieving its stated goal but through an unacceptable trade-off: for example, reducing network power consumption by shutting down multiple network towers and denying service to its customers.
For this reason, continuous validation frameworks are essential. For this, closed-loop digital twins should be used to test any AI-proposed change, with the RSG modeling the impact of these changes on all network KPIs in real time. The returned, simulated metrics are then used as immediate feedback that informs the AI’s next decision in a controlled, iterative loop.
One final oversight layer is still needed: an app validation engine. This monitors the long-term interaction between the application and the twin to ensure the AI avoids undesired trade-offs. By measuring primary objectives against guardrail KPIs like service quality and coverage, the validation engine quantifies these trade-offs and prevents the AI from diverging into inefficient or unstable states. Parallel models can simultaneously be run to provide benchmarks and score the application’s ultimate efficacy.
Stress testing the AI models
It’s vital to understand how the network and the AI RAN will perform when it encounters non-normal conditions, such as edge cases and attacks.
Any AI model must be robust, and a key function of a RAN digital twin is to prepare the AI applications for the unexpected.
Using the RSG to create what-if experiments within the sandbox environment of the digital twin helps identify weaknesses in the live network and improve resilience by showing how a system and any planned measures will cope in a specific situation. This might include events such as large-scale network congestion, hardware failures, or security threats, such as DoS and DDoS attacks, none of which can be tested at scale on a live network.
This same process should also be used to inform forward-looking development, such as cell tower location, or planning for the effects on RF propagation of a proposed/under-construction skyscraper. This brings us to the testing of 6G networks.
As was clear at the recent MWC Barcelona show, 2026 is set to be a core year for 6G, with field trials getting underway and standards being set. But there are several unknowns related to the technology. The RSG’s ability to undertake hypothesis-testing will therefore help bridge the gap between the lab and the field, creating trustworthy 6G propagation data to test 5G-6G spectrum sharing algorithms.
The technology will play a central role in the development of 6G infrastructure and de-risk the R&D process to accelerate 6G’s time to market.
A case study: energy-efficient 6G pre-deployment
As we’ve touched on above, digital twins and RSGs are being used to emulate and model a large array of situations, for example, balancing energy efficiency with QoE, or the implementation of Massive MIMO.
For example, the NVIDIA example referenced above is part of a novel approach leveraging AI models and agents to create a blueprint for intent-driven energy savings on 5G networks.
Figure 3. Architecture and key features of the Agentic AI blueprint for intent-driven 5G energy savings by VIAVI and NVIDIA. (Image: VIAVI and NVIDIA)
Another RSG use case that is worth highlighting comes from self-aware networks, with results confirming that this technique, which frees up wireless resources for data transmission, can improve system throughput.
One of the core elements in the study was on base station beam control. Traditionally, base stations select and control transmission beams based on network quality measurements reported by user devices. For the study, the twin was used to evaluate a novel methodology. Instead of relying on UE measurements and reports, the beam selection was made through a combination of AI-based predictions with network quality measurements obtained from the digital twin.
Figure 4. Simplified system setup for optimal beam selection evaluation based on the DOCOMO Self-awareness Network. (Image: VIAVI)
The joint study has confirmed that this methodology is able to provide a more optimal beam selection versus the standard approach. It also showed that, via the reduced radio control overhead offered by the technique, it was possible to achieve a 20% uplink throughput improvement. These results confirm that the use of digital twins and AI technologies can significantly reduce the frequency of UE network quality measurements and reporting, reducing control overhead.