If you’ve been wondering whether artificial intelligence is actually useful for research, or whether it’s just hype, the short answer is: it’s real, it’s here, and it’s moving fast.

We recently sat down with Alok Sahu, Senior Research Software Engineer at the AI Competency Centre, University of Oxford, to talk through what’s changed, what’s possible, and what researchers should be doing about it.

Why are LLMs Relevant to Research and Why Now?

Researchers have always worked with considerable volumes of text, reading papers, writing them, taking notes and reviewing others’ work. So, tools that handle text more efficiently have always had direct relevance. What’s shifted dramatically in the last couple of years isn’t the idea of using AI, it’s the quality of what these models can actually now do. Where they once struggled beyond a simple sentence completion, they can now summarise multiple long documents, draft and debug code, and carry out complex multi-step tasks with limited supervision.

The Leap to Agentic AI

One of the most significant recent developments is agentic AI – models that don’t just generate text but take actions. They can run literature searches, extract data from research papers, search the web, and return results, all in one workflow. For researchers, this opens up possibilities that simply weren’t practical before.

AI Is Not Just for One Field

The applications span multiple industries and sectors. In agriculture, Alok discussed with us, how AI is being used to predict crop yields and monitor fields. In construction, it’s flagging unsafe worksite behaviour and running design simulations. In healthcare, specialist models are emerging that can assist with drug discovery and genetic analysis. Whatever your research domain, there’s likely something relevant already developing.

New to LLM’s and AI? Where to Start

Alok’s advice is straightforward: dive in. Try the tools, experiment with your prompts, and see how the outputs change. Beyond the general platforms like ChatGPT, Claude, and Gemini, it’s worth exploring whether specialist models exist in your own field. The pace of development means domain-specific tools are emerging all the time.

The gap between those engaging with AI and those who aren’t is widening. The good news is it’s not too late to start and there’s no better time than now.

Watch the full conversation on our YouTube channel.

Connect with Alok:  https://www.linkedin.com/in/alokkrsahu/ 

Follow the AI Competency Center: https://www.linkedin.com/groups/15690177/ 

SMALLER, CHEAPER & GREENER MOBILE NETWORK HARDWARE

Mobile networks are quietly one of the most energy-intensive parts of our digital infrastructure. The towers, antennas, and radio units that keep us connected consume enormous amounts of power – and much of that consumption comes not from inefficiency in the obvious sense, but from a fundamental design problem that the industry has largely accepted as unavoidable. That is, until now.

The Problem: Duplicated Hardware, Multiplied Costs

Modern mobile networks operate across multiple frequency bands simultaneously. A single base station might need to serve users on a mid-band 5G frequency while also handling a separate private network on a different band. The way this has traditionally been handled is straightforward, but wasteful: build a separate radio unit for each band.

Two bands. Two sets of hardware. Two sets of power draws. Two sets of components to manufacture, deploy, and maintain.

Multiply that across thousands of base stations and the cost and carbon implications become significant. It’s one of the reasons that energy expenditure is a major operational burden for mobile network operators – and a genuine obstacle to meeting net-zero targets.

The question the University of Sheffield’s Wireless Communication Systems team set out to answer through the YO-RAN project was simple: why are we building two of everything, and can we stop doing that?

The Solution: One Radio Unit, Two Bands, a Single Shared Architecture

Demonstrated at Mobile World Congress 2026 in Barcelona, the Dual Band Open-RAN radio unit represents a significant step towards a leaner and more efficient model for mobile network hardware.

The core innovation is architectural. Rather than pairing two independent radio chains – each with its own digital-to-analogue converter (DAC), analogue-to-digital converter (ADC), and RF components – the Sheffield team has developed a radio unit in which two frequency bands share a single RF chain. Specifically, the prototype operates across the n78 (3.5 GHz) and n77-upper (4.005 GHz) frequency bands, which together cover both public 5G networks and private enterprise deployments.

The digital front-end (DFE) is built on the AMD/Xilinx ZCU670 RFSoC platform, a high-performance reconfigurable system-on-chip that handles two component carriers simultaneously. On the analogue side, a custom-designed RF front-end board manages both bands through shared circuitry – a single low-noise amplifier (LNA) for the receiver, and a shared transmit path enhanced with digital predistortion (DPD).

That last technique is worth pausing on. Power amplifiers are most efficient when operating close to their maximum output – but pushing them too hard distorts the signal. DPD corrects for that distortion digitally, allowing the amplifier to run at a more efficient operating point without sacrificing signal quality. It’s a meaningful contributor to reducing the overall power consumption of the unit. The result: fewer components, lower power draw, and a smaller physical footprint – without sacrificing performance.

Validated Against Industry Standards

Innovation in radio hardware only exists so far without interoperability. Open RAN – the disaggregated, vendor-neutral architecture that is reshaping how mobile networks are built – requires radio units to communicate with distributed units (DUs) from different suppliers using standardised interfaces.

The Sheffield prototype has been validated using the Keysight S5040A Distributed Unit Emulator, demonstrating compliance with the O-RAN 7.2x fronthaul interface specification. End-to-end testing confirmed that the dual band unit can exchange 5G NR signals cleanly across both frequency bands, with fronthaul synchronisation and RF chain performance verified across the full signal path.

This matters for the commercialisation pathway. A radio unit that works in a lab is one thing; a radio unit that can plug into a real Open-RAN ecosystem and interoperate with third-party DUs is quite another.

“The opportunity to present our neutral host, dual band, O-RAN radio unit technology at MWC was immensely valuable. The realisation of two concurrent, independent frequency bands on a unique single RF architecture yields significant cost and energy consumption savings. MWC and the HASC Hub provided the ideal environment to showcase this advanced radio technology, supporting our pathway to commercialisation and exploitation.”

~ Professor Timothy O’Farrell FREng,

Why This Matters Beyond the Lab

The implications of this work extend well beyond hardware engineering.

For network operators, a single compact radio unit covering two bands means lower capital expenditure on hardware, reduced installation complexity, and lower energy bills over the lifetime of the deployment. For operators running neutral-host models – where a single piece of infrastructure serves multiple operators or use cases – the ability to allocate the two bands independently is particularly powerful.

For vendors and the Open-RAN supply chain, simplified radio designs with fewer duplicated components reduce manufacturing costs and open up opportunities to deliver more competitive, sustainable products into a market that is increasingly cost-sensitive.

For society and the environment, the arithmetic is straightforward: lower-power radio units, deployed at scale across national and international networks, contribute significantly to reducing the energy footprint of mobile connectivity. At a time when both governments and operators are under pressure to demonstrate credible progress toward net-zero, such hardware-level efficiency gains matter.

The YO-RAN project, funded by DSIT, is one of a portfolio of research programmes housed within Sheffield’s National 6G Radio Systems Facility, a £2.4m testbed that provides one of the UK’s most capable platforms for physical layer 6G research.

What Comes Next

The prototype demonstrated at MWC is an early-stage proof of concept designed to establish that the architecture works and to inform the next phase of design refinement. Characterisation results from the prototype are now being used to optimise the hardware toward higher-TRL versions with improved RF performance and tighter integration.

The team is actively seeking research collaborators and industry partners to help accelerate that journey from prototype to deployment. If you’re working in network infrastructure, Open-RAN, spectrum policy, or sustainable connectivity and want to explore what a partnership might look like, get in touch.

6gfacility@sheffield.ac.uk  

The YO-RAN project is funded by the Department for Science, Innovation and Technology (DSIT). The National 6G Radio Systems Facility is funded by EPSRC. The work was presented at Mobile World Congress 2026, Barcelona, with support from the Hub for Access to the Spectrum Community (HASC).

HASC Research Pillar: C2 Adaptivity | Imperial College London

What if a wireless network could learn to serve you better – without needing to ask where you are, what device you’re holding, or which way you’re facing? That’s exactly the question a team at Imperial College London set out to answer.

Led by Professor Kin Leung and Dr. Nancy Nayak, this research sits at the heart of HASC’s Adaptivity pillar – a challenge focused on building networks that can intelligently respond to changing conditions in real time. Their approach: replace the rigid, standards-dependent methods that underpin today’s mobile networks with something far more flexible. An AI that learns.

“We are not just jumping on to the band wagon of AI – we have developed AI-based wireless communication technologies which are fundamentally different from the conventional technologies used today.” ~ Professor Kin Leung, Imperial College London

“By designing wireless technologies with AI, we move beyond dependence on channel state information—unlocking faster development cycles and enabling networks to adapt to user needs and dynamic environments, unconstrained by standards.” ~ Dr. Nancy Nayak, Imperial College London

Why We’re Experimenting

The demand on wireless networks is accelerating. More users, more devices, more data – and the greater the expectation. People want and expect seamless connectivity everywhere, from city centres to crowded stadiums. Meeting those demands with today’s tools is increasingly difficult, and the gap between what users want and what networks can reliably deliver is widening.

The team at Imperial saw an opportunity to fundamentally rethink how networks allocate their resources. Rather than optimising within the constraints of existing standards, they asked: what if the AI could simply observe a network in action and learn the best way to manage it – without any of the traditional scaffolding?

The Industry Challenge

At the core of most modern wireless systems is something called Channel State Information, or CSI. This is data that describes the communication channel between a base station and a user’s device – think of it as the network’s way of understanding who’s where and how best to reach them.

The problem is that collecting and communicating CSI comes at a cost. For complex antenna systems (the kind increasingly used in 5G and future 6G networks) the overhead is significant. And because CSI exchange relies on standardised protocols, any improvement to the process must go through the slow, uncertain machinery of the international standards process, which can take years.

Add to this, the computational burden of recalculating antenna configurations as users move, and the challenge becomes clear: current approaches don’t scale well to the networks of tomorrow.

Our AI-Enabled Technology Innovation

The Imperial team has developed a Reinforcement Learning (RL) technique that sidesteps these constraints entirely. Their system allocates wireless resources (directing antenna beams and managing radio resources) without relying on CSI, user location data, or any information obtained through standardised protocols.

Instead, it learns. Using only the radio measurements already available within the network, the AI explores different resource-allocation strategies, receives feedback on what works, and progressively improves – just like a person learning a new skill through practice rather than instruction.

The results are striking. In testing, the approach achieves performance within 6% of the theoretical optimum for CSI-based methods – without any of the associated overhead. The system can also co-exist with conventional CSI-based approaches used by neighbouring base stations, making it compatible with real-world deployment scenarios.

Crucially, because the technology is standards-agnostic, operators don’t need to wait for industry alignment before deploying it. New antenna technologies and networking innovations can be adopted as soon as they’re ready.

The Impact of CSI-Free Networks 

The implications reach well beyond the lab. For network operators, this approach reduces complexity, lowers the cost of upgrades, and dramatically shortens the path from innovation to deployment. For users, it means more reliable connections in the places that matter most: busy transport hubs, large events, dense urban environments.

Looking further ahead, the team is building a prototype for cellular network settings and exploring civilian and defence applications – with defence interest from Defence Science and Technology Laboratory (Dstl) and the Royal Air Force. The commercial opportunity is significant, with the short-term market for private 5G and enterprise networks estimated at $20–200 million, and long-term integration into baseband units pointing to a $0.5–1.5 billion opportunities. The global baseband unit market is ~$5–7B today and growing beyond $10B; as a core signal processing function (~5–20% share), resource allocations including beamforming underpin this serviceable segment  [1, 2, 3].

This is a project that doesn’t just improve a single component of how wireless networks function. It reframes how they can be built, deployed, and improved – putting adaptivity, rather than standardisation, at the centre.

Interested in this research or the path to commercialisation?  

We’re currently keeping our IP confidential while we finish our patent filings. At the same time, we’ve started building a live demo to show the tech in action. This demo is a key part of our spinout plan, as it will give investors a clear look at the value the new AI wireless technology brings to network users.

If you would like to get in touch with this project, please contact us here. 

Meet ORLANDO – the AI-powered xApp making wireless networks smarter, safer, and more resilient

Picture a busy summer festival. Thousands of people are live-streaming, uploading, chatting and scrolling – and then an emergency happens. First responders try to make contact, but the network is overloaded. They can’t get through.

Network congestion isn’t just frustrating. In critical moments, it can be dangerous. This is exactly the problem that researchers at the University of York, in collaboration with Imperial College London, are working to solve – through a project called ORLANDO (O-RAN intelligent adaptive Load blaNcing and efficiency in highly Dense deplOyments).

Why We’re Experimenting

As we move towards 6G, networks must serve more devices, more applications, and more people – all at once, and in dense environments. The ability to balance traffic intelligently and adapt in real time, is no longer a nice-to-have. It’s essential.

ORLANDO sits within the HASC Core Challenge on Adaptivity: a research programme dedicated to building networks that can think, learn, and self-optimise. The project is investigating how AI and machine learning (ML) can be embedded directly into Open RAN (O-RAN) architecture to manage load balancing at scale.

The Industry Challenge

Dense O-RAN networks struggle to maintain performance under uneven traffic loads. When demand spikes (at a concert, a sports stadium, or a major public event for example) spectrum is wasted, latency climbs, and quality of service degrades. Traditional, rule-based systems simply can’t respond fast enough to keep up with the unpredictability of real-world demand.

Our Technology Innovation

ORLANDO is an AI-powered xApp – a software application that runs within the O-RAN Intelligent Controller (RIC) – designed to perform real-time network slicing and intelligent load balancing. Using a digital twin integrated with live network data and precise access point locations, the system simulates and optimises performance across different user load scenarios before applying changes to the live network.

The team is also training a traffic prediction ML model on real user movement and traffic data – and fine-tuning a generative AI to produce synthetic data for new or unforeseen environments. This makes ORLANDO not just reactive, but predictive.

The Impact of the ORLANDO Project

Already tested on the York O-RAN testbed – and due to be trialled in a real-world deployment in Blackpool – ORLANDO is designed to deliver tangible benefits: dynamic traffic prioritisation, optimised Quality of Service (QoS), and reliable connectivity for critical services, even at peak demand.

For citizens, this means fewer dropped calls, less buffering, and the confidence that emergency services will always be able to get through – even in a crowd. For industry, it signals a new era of intelligent, self-optimising networks – built not just for today’s demands, but for whatever comes next.

ORLANDO Team Showcase

“In ORLANDO we move a step forward towards the AI-Native networks where we learn the user behaviour and train a scalable ML to predict network load and allocate network resources accordingly.” ~ Dr. Hamed Ahmadi, Reader in Digital Engineering

To connect directly with the ORLANDO researchers, register your here. 

See ORLANDO – The Intelligent Load Balancing xApp from University of York

Links, Papers & Further Resources    

• Scalable machine learning-based approaches for energy saving in densely deployed Open RAN – https://arxiv.org/pdf/2604.00201
• Unlocking Efficient Connectivity: Advanced AI-Driven Planning for Multi-Hop IAB Networks – https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11426753
• Interpretable Attention-Based Multi-Agent PPO for Latency Spike Resolution in 6G RAN Slicing – https://arxiv.org/pdf/2602.11076
• Green O-RAN Operation: a Modern ML-Driven Network Energy Consumption Optimisation – https://arxiv.org/pdf/2512.07006

Connect with Hamed Ahmadi

• https://www.linkedin.com/in/hamed-ahmadi-24504b36/
• https://www.itwins-lab.uk/
• https://scholar.google.com/citations?hl=en&user=lEcMAWwAAAAJ&view_op=list_works&sortby=pubdate

The security of communication networks is usually discussed in cryptographic terms, such as asymmetric cryptography and symmetric encryption. However, these cannot easily be applied to the billions of low-cost, limited-capability Internet of Things (IoT) devices coming online. This presents a major security risk, potentially giving malicious actors a back door into our private lives.

Yet every wireless connection – from a baby monitor to a satellite link – fundamentally depends on electromagnetic signals travelling through the physical environment. This ‘physical layer’ of radio frequency (RF) communication is becoming a promising new frontier for providing security across the wireless network.

What is RF physical layer security?

Physical layer security (PLS) describes security protocols applied to the very lowest layer of communications, and which exploit the inherent physical properties of channels. For RF PLS, this means harnessing characteristics such as noise, interference, and channel fading.

Unlike cryptography, which relies on the difficulty of solving a mathematical problem and hiding algorithms, PLS exploits physics – in particular the randomness of the wireless environment and the unique behaviour of real-world hardware.

Every wireless link has its own, location-dependent behaviour. Walls, furniture, other objects, and people moving around all cause signals to reflect, scatter and fade in ways that are unique to the positions of the transmitter and receiver.

Two legitimate devices in fixed positions ‘see’ the same rapidly changing channel between them and can treat it as a shared source of randomness for generating keys or authenticating each other. An eavesdropper in a different place, even a few wavelengths away (over 10 cm for Wi-Fi), experiences a different channel pattern, so cannot easily reproduce the same measurements or derive the same secret information. In this way, the physical environment itself becomes part of the security mechanism.

Why does physical layer security matter now?

Traditional cybersecurity tools have worked reasonably well for laptops, servers, and smartphones. These devices have powerful processors and large memories that can support heavyweight cryptographic algorithms.

The picture is different for the tens of billions of IoT devices now being deployed in homes, factories, cities, and elsewhere. Many of these sensors and embedded devices are:

• extremely low cost,
• battery-powered, with tight energy constraints,
• designed for ‘install and forget’ deployments in unattended environments.

In these settings, strong cryptography can be difficult to deploy or maintain. PLS, on the other hand, can be implemented with lightweight signal processing, can operate continuously, and be applied to legacy devices. Ultimately, PLS is not a replacement for cryptography, but a way to raise the baseline for these vulnerable endpoints – the ‘Achilles heel’ of the network.

HASC leadership

Researchers within the Hub for All Spectrum Connectivity (HASC) are leading the charge in three PLS areas in particular:

1. Physical-layer authentication

Physical-layer authentication uses channel measurements – for example, detailed channel state information – to create an ‘environmental fingerprint’ for a device. If a device suddenly appears with a channel profile that does not match its historical pattern, this could indicate spoofing or impersonation. HASC researchers have developed a practical deep learning-based physical-layer authentication scheme that can cope with mobile, time-varying wireless channels.

By training a neural network on a mix of synthetic and real Wi-Fi channel data, they reduced the amount of field measurements needed while still learning to distinguish between genuine devices and impostors. This shows that channel-based authentication can remain reliable even when users and devices are moving.

2. Wireless key generation

In wireless key generation, two legitimate devices observe their shared channel and run signal-processing algorithms to extract a shared secret key from the common randomness in the channel. Because an eavesdropper does not see the same channel, they cannot easily derive the same key.

Much early work in this area has been based on a single link between a pair of users. However, in real IoT networks, a central node typically interacts with many devices. HASC researchers have addressed this challenge by designing an efficient multi-user key generation protocol based on Wi-Fi 6 orthogonal frequency-division multiple access (OFDMA), which allows a central node to split the spectrum into many subchannels and talk to multiple users at the same time. The protocol uses OFDMA capability so that the access point can probe and harvest randomness from several user channels in parallel.

HASC researchers have also conducted extensive experimental work on key generation for Wi-Fi and long-range IoT technologies, opening up additional security options for future multi-technology 6G systems.

3. Radio frequency fingerprinting (RFF)

Every RF transmitter, even when manufactured to the same specifications, has subtle hardware imperfections in its amplifiers, oscillators, mixers, and other components. These imperfections imprint a unique, repeatable signature on the transmitted waveform.

Machine learning-based RF fingerprint identification (RFFI) can learn and recognise these signatures, enabling device-level authentication at the physical layer. Deep learning models, originally developed for image and speech recognition, are particularly well-suited to learning these patterns and classifying multiple signals.

HASC researchers have carried out comprehensive studies on exploring how RF fingerprinting could add an additional layer of security to networks the moment a device starts transmit information, without relying solely on user credentials. Their work ranges from LoRa, Wi-Fi, Bluetooth low energy to LTE.

HASC researchers at the University of Liverpool, Heriot-Watt University, and Queen’s University Belfast have demonstrated real-time RF fingerprinting using off-the-shelf Wi-Fi USB dongles, identifying which of multiple dongles sent a given packet based purely on their RF ‘fingerprints.’

Another HASC research focus is developing methods to ensure that RFF is applicable in the real world, and not just controlled, laboratory settings. For instance, a recent HASC study showed that advanced AI models, such as denoise diffusion models, can first suppress channel noise and effectively ‘amplify’ unique RF fingerprints, making them detectable and reliable even in noisy, real-world wireless environments.

Challenges and open questions

PLS could play a key role in supporting the UK Government’s Secure by Design agenda, and recently-introduced laws that require device manufacturers to implement minimum security standards.

But despite the rapid progress in PLS research, several challenges need to be addressed before this can become routine in commercial products:

• Standardisation and integration: PLS techniques are not yet widely embedded in mainstream standards. Integrating them without disrupting existing communication systems is a major research and engineering challenge.
• Scalability: Many experimental studies focus on tens of devices, but real deployments may involve thousands or millions of devices. Ensuring that key generation, authentication, and RF fingerprinting scale robustly remains an open question – one that requires more realistic testbeds and larger trials to answer.
• Access to physical-layer data: Techniques like channel-based authentication often require detailed channel state information that is not typically exposed by commercial RF chipsets. Closer collaboration with chipset vendors and equipment manufacturers will be important to unlock these capabilities at scale.
• All-spectrum connectivity: As we move towards a vision of multi-tech networks spanning RF, millimetre-wave, terahertz and optical links, this will create new opportunities to harness the diversity of channels for security – but also new opportunities for attack that need to be understood and managed.

PLS will never replace cryptography, but it offers a promising pathway to making wireless systems more resilient, adaptable, and trustworthy. By leveraging the underlying physics of the radio environment, HASC research is helping to make future networks inherently more secure from the moment a device connects.

With special thanks to Dr Junqing Zhang, University of Liverpool