Can machine learning solve the challenges of non-terrestrial networks?
Satellite and telephone connections are becoming increasingly popular. The service now provides direct satellite connectivity to regular smartphones, extending coverage to locations where traditional cellular infrastructure does not exist. Granted, most of these services are text-only, but there’s no reason to think they won’t continue to evolve.
What about the prey? Those satellites don’t sit still. Satellite overhead is eliminated in minutes, and connections should seamlessly hop to ground infrastructure or another satellite unnoticed. Network management systems built for fixed towers were not designed with this type of constant movement in mind.
This is where AI could play a role in how non-terrestrial networks function. Coordinating handovers between fast-moving satellites and fixed terrestrial networks requires real-time decisions across variables that refuse to remain constant. AI could be the technology that can navigate these dynamic situations, unraveling the complex relationships between network parameters that overwhelm static, rule-based approaches.
Challenges of non-terrestrial networks
Non-terrestrial networks operate under very different constraints than traditional 5G and emerging 6G systems. The most obvious difference is the movement itself. While the satellites follow a continuous orbit, the cell phone towers remain in place.
The Doppler effect poses one of the bigger problems. When a satellite moves in position relative to a ground station or user device, the signal frequency can vary significantly. Anyone who has ever heard the pitch of a siren change as an ambulance passes by understands the basic physics, but in communication systems that rely on precise frequency coordination, this phenomenon causes serious disruption. Ground equipment and user devices must be compensated at all times.
Delays in propagation further complicate the problem. LEO satellites also experience delays that disrupt timing synchronization between network components. Every millisecond really matters to keep a call or video streaming, and the round trip distance to orbit introduces timing complexities never encountered in terrestrial networks.
The most difficult challenge may be the satellite coverage pattern itself. The trajectory is predictable, but the coverage gap changes continuously. One satellite may fall below the horizon and a replacement satellite may not be in an ideal position to pick up the slack. Static algorithms are hampered at these transitions because the optimal decision depends on too many interacting factors, such as satellite position, user position, network congestion, atmospheric conditions, and ground infrastructure conditions.
AI-powered handover management
AI tackles handover complexity through predictive algorithms that predict coverage loss before it occurs. Rather than scrambling the signal after it drops, these systems digest orbital data, ground network topology, and real-time conditions to determine exactly when handoff occurs and where the connection will reach next.
Pre-positioning is one of the major advantages. The AI system will begin establishing links with neighboring cells or satellites before the current connection degrades. By the time the actual handover occurs, the transition infrastructure is already in place, reducing the gap that would mean dropped calls and session interruptions. From the user’s perspective, the connection continues without any problems or knowledge that anything has happened.
These algorithms become smarter through pattern recognition. Although orbital mechanics follows known physics, real-world performance depends on a myriad of additional factors, including weather, terrain, building interference, and traffic patterns. Machine learning identifies the variables that most influence handover success in a given situation and adjusts decisions accordingly. A system that initially fails to handover in a mountainous region will eventually be able to learn the specific characteristics of its environment.
However, in real deployment, constraints arise that cannot be fully captured by experimental results. Most of the research in this area is experimental. Hardware limitations, tight latency budgets, and the need to operate across multiple communication standards all present challenges that cannot be fully predicted in controlled demonstrations.
AI in space
Handoff management is just one piece of a larger puzzle. AI performs several interconnected functions to keep the satellite network operational.
Signal processing and spectrum management are important. Satellites need to share frequencies with ground users without causing interference, which requires dynamic spectrum adjustment. The AI-based system identifies available frequencies, handles real-time demodulation, and prevents satellite transmissions from interfering with terrestrial communications, while protecting the satellite link from terrestrial interference.
Resource allocation becomes quite complex when both satellite and terrestrial infrastructure are involved. Network slicing, or dividing bandwidth and computing resources based on demand, requires AI systems that can respond to changing patterns from minute to minute and dynamically shift capacity to where it’s actually needed.
Beam management also benefits from predictive capabilities. Satellites communicate through focused beams, so optimizing beam direction means predicting where users will travel. AI systems predict movement patterns and adjust location to maintain strong connections with mobile users.
Anomaly detection solves problems. AI systems can spot performance degradation or new failures before users notice any anomalies, flagging issues and intervening before calls are dropped or connections become slow.
Implementation obstacles and scalability
Current satellite hardware places severe limits on what AI can actually accomplish in orbit. Chipsets and FPGA platforms face power and thermal constraints that limit algorithmic complexity. Advanced AI models require processing power that satellite payloads cannot currently provide, forcing engineers to make trade-offs between accuracy and computational feasibility.
Propagation delays often require AI systems to calculate decisions in advance, rather than reacting on the fly. This approach works well in predictable scenarios, but becomes difficult when conditions change unexpectedly.
Questions about scalability remain largely unanswered, at least for now. Research has shown promising results in controlled testbeds, but supporting millions of concurrent users across global mega-constellations raises issues that no one has fully resolved. Regulatory fragmentation slows everything down. Different countries and standards bodies, such as 3GPP for 5G today and various 6G working groups, are moving at different speeds and sometimes in contradictory directions. Systems need to be designed with flexibility in mind, which further complicates engineering.
Testing presents unique challenges because real satellite conditions resist ground-based simulation. Research facilities like 6GSPACELab use commercial hardware such as FPGA platforms and AI chipsets to build practical testbeds to validate signal processing techniques in both ground and orbit scenarios. The current experimental system runs a 112 Gbps interconnect between the AI processor and the RF system for real-time decision making. However, extensive field testing is still inevitable before commercial deployment, and that testing is neither quick nor cheap.
The future of 6G and satellite AI
The research consensus is that AI will be fundamental to the satellite layer of 6G, but that deployment will occur in stages rather than in one dramatic leap.
Short-term work over the next two to four years will focus on improving spectrum sharing, basic mobility optimization, and handover between satellite and ground systems. These goals are achievable with existing technology and establish the basis for more ambitious applications.
Medium-term advances in 4 to 7 years are likely to result in 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. This indicates that serious organizational resources are flowing in this direction.
It is already being put into practical use in certain fields. Machine-to-machine use cases require continuous connectivity across remote locations where terrestrial communication coverage does not make economic sense. Self-driving cars are another application, where satellites provide backup navigation, map updates, traffic data, and emergency services beyond cell phone coverage. IoT sensors, already in the billions and geographically dispersed, benefit from satellite connectivity managed by AI systems optimized for degraded devices.
Expectations that AI will simply solve the handoff problem require some skepticism. While it is clearly an enabling technology, success will equally depend on infrastructure support, regulatory alignment, and continued advances in both hardware and algorithms. Between laboratory demonstrations and seamless global coverage, there are challenges that only emerge once deployment begins.
