Can machine learning solve the challenges of non-terrestrial networks?
Satellite-to-phone connectivity is becoming a whole lot more popular. Services are now delivering direct satellite connections to regular smartphones, extending coverage into places where traditional cell infrastructure simply doesn’t exist. Sure, most of those services are text-only, but there’s no reason to assume they won’t get more advanced.
The catch? Those satellites don’t sit still. A satellite overhead is gone in minutes, and your connection has to hop seamlessly to ground infrastructure or another satellite without you ever noticing. Network management systems built for stationary towers weren’t architected for this kind of constant motion.
This is where AI could play a role in how non-terrestrial networks work. Orchestrating handovers between fast-moving satellites and fixed ground networks demands real-time decisions across variables that refuse to stay constant. AI may well be the technology capable of navigating these dynamic conditions — teasing out complex relationships among network parameters that would overwhelm any static rule-based approach.
The challenge of non-terrestrial networks
Non-Terrestrial Networks operate under constraints that look nothing like conventional 5G or emerging 6G systems. The most glaring difference is movement itself — cell towers stay planted in place while satellites trace continuous orbital arcs.
The Doppler effect creates one of the bigger problems. As satellites shift position relative to ground stations or user devices, signal frequencies drift, sometimes substantially. Anyone who’s heard an ambulance siren pitch-shift as it drives past understands the basic physics, but in communications systems dependent on precise frequency coordination, this phenomenon causes serious disruption. Ground equipment and user devices have to compensate constantly.
Propagation delays compound matters further. Even LEO satellites introduce latency that throws off timing synchronization between network components. For maintaining calls or streaming video, milliseconds actually matter, and the round-trip distance to orbit creates timing complications terrestrial networks never encounter.
The most difficult challenge might be satellite coverage patterns themselves. Orbits are predictable, yet coverage gaps shift continuously. One satellite drops below the horizon while its replacement may not be ideally positioned to pick up the slack. Static algorithms choke on these transitions because optimal decisions depend on too many interacting factors — satellite positions, user locations, network congestion, atmospheric conditions, ground infrastructure status, and so on.
AI-driven handover management
AI tackles handover complexity through predictive algorithms that see coverage loss coming before it arrives. Instead of scrambling after a signal drops, these systems digest orbital data, ground network topology, and real-time conditions to determine precisely when handoffs should happen and where connections should land next.
Pre-positioning is one of the major advantages. AI systems start establishing links with neighboring cells or satellites before current connections degrade. By the time actual handover occurs, the transition infrastructure is already in place, shrinking the gap that would otherwise mean dropped calls or interrupted sessions. From the user’s perspective, the connection just continues, with no hiccup or awareness that anything happened.
These algorithms get smarter through pattern recognition. Orbital mechanics follow known physics, but real-world performance hinges on countless additional factors: weather, terrain, building interference, congestion patterns. Machine learning identifies which variables most influence handover success in specific contexts and adjusts decision-making accordingly. A system that initially fumbled handovers in mountainous regions can eventually learn the particular characteristics of that environment.
Real-world deployment introduces constraints that lab results don’t fully capture, though. Most research in this space remains experimental. Hardware limitations, strict latency budgets, and the need to work across multiple telecommunications standards all present challenges that controlled demonstrations can’t entirely anticipate.
AI in space
Handover management represents just one piece of a larger puzzle. AI performs several interconnected functions that keep satellite networks operational.
Signal processing and spectrum management are critical. Satellites must share frequencies with terrestrial users without causing interference, requiring dynamic spectrum coordination. AI-based systems identify available frequencies, handle real-time demodulation, and prevent satellite transmissions from stepping on terrestrial communications, while also protecting satellite links from ground-based interference.
Resource allocation grows considerably more complex when satellite and ground infrastructure both enter the equation. Network slicing, or carving up bandwidth and compute resources based on demand, requires AI systems that can dynamically shift capacity where it’s actually needed, responding to patterns that change minute by minute.
Beam management also benefits from prediction capabilities. Satellites communicate through focused beams, and optimizing beam direction means anticipating where users will move. AI systems project movement patterns and adjust positioning to maintain strong connections even with mobile users.
Anomaly detection rounds things out. AI systems can spot performance degradation or emerging faults before users notice anything wrong, flagging problems for intervention before calls drop or connections slow.
Implementation obstacles and scalability
Current satellite hardware places hard limits on what AI can actually accomplish in orbit. Chipsets and FPGA platforms face power and thermal constraints that bound algorithmic complexity. Sophisticated AI models demand processing power that satellite payloads can’t currently provide, forcing engineers into tradeoffs between accuracy and computational feasibility.
Propagation delays mean AI systems frequently need to pre-compute decisions rather than react in the moment. This approach works adequately for predictable scenarios but struggles when conditions shift unexpectedly.
Scalability questions remain largely unanswered, at least for now. Research shows promising results in controlled testbeds, but supporting millions of simultaneous users across global mega-constellations introduces problems nobody has fully solved. Regulatory fragmentation slows everything down. Different countries and standards bodies, including 3GPP for current 5G, and various 6G working groups, move at their own speeds and sometimes in conflicting directions. Systems must be designed for flexibility, which adds engineering complexity.
Testing poses unique difficulties since real satellite conditions resist ground-based simulation. Research facilities like 6GSPACELab are building practical testbeds using commercial hardware, like FPGA platforms and AI chipsets, to validate signal processing techniques for both ground and orbital scenarios. Current experimental systems run 112 Gbps interconnects between AI processors and RF systems for real-time decision-making. But extensive field testing remains unavoidable before commercial deployment, and that testing is neither quick nor cheap.
The future of 6G and satellite AI
The research consensus holds that AI will be fundamental to 6G’s satellite layer, though deployment will happen incrementally rather than in one dramatic leap.
Near-term work over the next two to four years focuses on spectrum sharing between satellite and terrestrial systems, basic mobility optimization, and handover refinement. These goals are achievable with existing technology and establish groundwork for more ambitious applications.
Medium-term advances in the four to seven year window will likely bring more sophisticated resource allocation, improved indoor/outdoor handover handling, and integration with terrestrial edge computing for distributed inference. The European Space Agency is funding research into AI-optimized satellite systems through its 6G Satellite Precursor initiative — a signal that serious institutional resources are flowing toward this direction.
Practical applications are already taking shape in specific domains. Machine-to-machine use cases equire continuous connectivity across remote areas where terrestrial coverage makes no economic sense. Autonomous vehicles represent another application, with satellites providing backup navigation, mapping updates, traffic data, and emergency services beyond cell coverage. IoT sensors, already numbering in the billions and scattered geographically, benefit from satellite connectivity managed by AI systems optimized for reduced-capability devices.
The expectation that AI will simply solve the handover problem deserves some skepticism. It’s clearly enabling technology, but success depends equally on supporting infrastructure, regulatory alignment, and ongoing advances in both hardware and algorithms. The gap between laboratory demonstrations and seamless global coverage involves problems that only emerge once deployment actually begins.
