Revolutionizing Future Connectivity: A Contemporary Survey on AI-empowered Satellite-based Non-Terrestrial Networks in 6G

Shadab Mahboob; Lingjia Liu

Source-linked AI summary

Revolutionizing Future Connectivity: A Contemporary Survey on AI-empowered Satellite-based Non-Terrestrial Networks in 6G

Shadab Mahboob, Lingjia Liu

arXiv:2303.01633v4Mar 2, 2023cs.NIeess.SP

TL;DR

Future 6G needs ubiquitous connectivity, but satellite-based NTNs introduce propagation, mobility, spectrum, allocation, and integration challenges. This survey reviews NTN and AI foundations, maps AI methods to NTN problems, and examines research directions, testbeds, software-defined integration, and practical deployment issues. It concludes that AI is promising for satellite-based NTN, while real deployment remains constrained by latency, onboard resources, security, and online implementation requirements.

Problem
Existing work does not comprehensively cover AI-enabled NTN-integrated 6G networks, their specific challenges, practical complications, and current implementation efforts.
Method
The paper surveys NTN architectures and challenges, AI approaches, relevant research, testbeds, software-defined integration efforts, and practical issues for satellite-based 6G NTNs.
Results
RL is primarily used for NTN optimization problems, while SL is used for estimation problems; UL approaches receive comparatively limited coverage.
Takeaways & Limitations
AI techniques offer a framework for addressing diverse satellite-based NTN challenges, but their use must account for real-time feedback, resource constraints, and deployment conditions.
Takeaways & Limitations
AI-driven NTN deployment is constrained by cost-limited onboard capabilities, time-varying satellite networks, long propagation delays, and security and online-implementation challenges.

Abstract

from arXiv · show

Non-Terrestrial Networks (NTN) are expected to be a critical component of 6th Generation (6G) networks, providing ubiquitous, continuous, and scalable services. Satellites emerge as the primary enabler for NTN, leveraging their extensive coverage, stable orbits, scalability, and adherence to international regulations. However, satellite-based NTN presents unique challenges, including long propagation delay, high Doppler shift, frequent handovers, spectrum sharing complexities, and intricate beam and resource allocation, among others. The integration of NTNs into existing terrestrial networks in 6G introduces a range of novel challenges, including task offloading, network routing, network slicing, and many more. To tackle all these obstacles, this paper proposes Artificial Intelligence (AI) as a promising solution, harnessing its ability to capture intricate correlations among diverse network parameters. We begin by providing a comprehensive background on NTN and AI, highlighting the potential of AI techniques in addressing various NTN challenges. Next, we present an overview of existing works, emphasizing AI as an enabling tool for satellite-based NTN, and explore potential research directions. Furthermore, we discuss ongoing research efforts that aim to enable AI in satellite-based NTN through software-defined implementations, while also discussing the associated challenges. Finally, we conclude by providing insights and recommendations for enabling AI-driven satellite-based NTN in future 6G networks.

I. INTRODUCTION

6G requires ubiquitous, high-performance connectivity that terrestrial-only infrastructures cannot guarantee, motivating satellite-based NTN integration despite substantial propagation, mobility, and coordination challenges. This survey organizes NTN and AI foundations, reviews AI-based solutions and implementation efforts, and identifies practical research directions for integrated 6G networks.

I. INTRODUCTION: Terrestrial-only infrastructures cannot guarantee ubiquitous connectivity, while NTNs use space- and air-borne platforms to extend network coverage.The survey focuses on satellite-based NTN because satellites offer ubiquitous coverage, predictable trajectories, and scalability.
I. INTRODUCTION: Satellite-based NTN faces higher propagation delay, substantial Doppler shift, modified handover and paging requirements, and interference-management challenges during terrestrial integration.The integrated environment also requires efficient computing, routing, and slicing algorithms to meet 6G KPI requirements.
I. INTRODUCTION: The survey addresses a literature gap by systematically reviewing AI methods for NTN challenges, related research progress, testbeds, software-defined integration, and practical complications.It also provides background on NTN and AI, future research scopes, and recommendations for satellite-based NTNs in 6G.
I. INTRODUCTION: The paper proceeds from NTN architectures, platforms, use cases, and challenges to AI approaches, related surveys, and AI solutions for NTN challenges.Its organization connects the background discussion with research trends that combine identified challenges and AI techniques.

2) Air-borne platforms:

Air-borne NTN platforms include HAPS, while satellites remain central because of their broader coverage and operational advantages. NTN applications span high-throughput, low-latency, extreme-coverage, and massive-connectivity requirements, often alongside terrestrial networks.

2) Air-borne platforms:: High Altitude Platform Systems (HAPS) are air-borne wireless communication platforms, including airships, balloons, and airplanes, typically operating around 20 km altitude.HAPS have smaller propagation delay than space-borne platforms but face stabilization and refueling challenges.
2) Air-borne platforms:: Satellites are prioritized over airborne platforms because they provide global coverage, stable and predictable orbits, scalability, and established international regulations.Accordingly, the paper primarily studies satellite-enabled NTNs in 6G.
2) Air-borne platforms:: NTNs support application categories combining eMBB with mMTC, URLLC, or extreme-coverage requirements, including digital twins, V2X, intelligent transport, and ubiquitous connectivity.The passage identifies eMBB, mMTC, and URLLC as performance-based application categories.
2) Air-borne platforms:: Satellites can reach underserved locations and support disaster scenarios, while LEO systems offer relatively low latency for some real-time applications.The paper notes that the most latency-sensitive 5G URLLC and 6G use cases may require conjunction with terrestrial networks.

C. General Architecture

Satellite-based NTN architectures connect user equipment and terrestrial networks through satellite payloads, gateways, feeder links, and core-network backhaul. Their long transmission distances create substantial propagation delay and path loss, shaped by satellite altitude and carrier frequency.

Architecture: Transparent payloads perform RF filtering, frequency conversion, and amplification, while regenerative payloads add onboard processing after modulation and coding.
Architecture: Satellite-based NTN uses satellites, gateways, user equipment, and feeder links to connect terrestrial networks and public data networks.Transparent payloads relay RF-processed signals, whereas regenerative payloads process signals onboard and operate like base stations.
Propagation characteristics: NTN free-space path loss is approximately 60-120 dB and depends on distance and signal frequency.Atmospheric-gas, rain, fog, and scintillation attenuation further affect propagation loss, particularly at higher frequencies.
Satellite operation: Satellite mobility and propagation conditions differ across GEO and NGEO systems, requiring distinct considerations for network operation.GEO satellites appear relatively static with respect to the ground, whereas NGEO scenarios differ substantially.

5) Coverage Area:

Satellite coverage enables connectivity over very large and remote areas, but large footprints, NGEO mobility, and propagation conditions complicate timing, handovers, synchronization, and resource management. Integrating NTN with terrestrial systems also creates challenges for spectrum sharing, offloading, routing, and slicing.

5) Coverage Area:: Large satellite beam footprints provide ubiquitous coverage over remote and isolated areas, exceeding the coverage areas of terrestrial counterparts.Users at different positions within a large cell experience different delays, requiring modified timing and synchronization procedures.
5) Coverage Area:: NGEO satellites remain visible to a terrestrial user equipment for only several minutes, causing multiple handovers even when the user is stationary.Spot-beam operation can require additional beam handovers because individual beams cover smaller areas.
5) Coverage Area:: NGEO mobility introduces Doppler shifts that create frequency offsets, disrupting frequency synchronization and potentially causing interference.
5) Coverage Area:: High path loss and limited spectrum make NTN spectrum and power allocation complex, while terrestrial users create co-channel interference in target bands.S-band and Ka-band are identified as target NTN bands with existing or emerging terrestrial communication use.
5) Coverage Area:: Long propagation delay makes timing-advance management and computational offloading difficult in integrated terrestrial-NTN networks.The integration also raises open issues in network routing and network slicing.
5) Coverage Area:: Satellite-based NTNs offer ubiquitous connectivity, service continuity, and extreme reliability while exposing AI-relevant challenges from mobility, distance, spectrum sharing, and propagation loss.

III. AI AND ITS RELEVANCE TO NTN CHALLENGES

The paper focuses on learning-based AI because satellite-based NTN has a complex, time-varying topology that is difficult to address with predefined rules. It introduces ML workflows involving model initialization, training, testing, and feature extraction, while motivating neural networks for nonlinear high-dimensional problems.

III. AI AND ITS RELEVANCE TO NTN CHALLENGES: The paper emphasizes learning-based AI for NTN because its topology is extremely complex and time-varying, making rule-based approaches less feasible.
III. AI AND ITS RELEVANCE TO NTN CHALLENGES: A generic ML model comprises pre-training, training, and testing phases, with learning-strategy selection and parameter initialization occurring before training.
III. AI AND ITS RELEVANCE TO NTN CHALLENGES: Training uses preprocessed data and selected features to capture input correlations, evaluates model outputs, and updates the model using evaluator feedback.
III. AI AND ITS RELEVANCE TO NTN CHALLENGES: Testing evaluates a trained model with processed testing data, whereas offline and online learning differ in whether training data is generated all at once or incrementally.
III. AI AND ITS RELEVANCE TO NTN CHALLENGES: Generic ML models can struggle with hundreds of parameters and nonlinear input-output relationships, motivating neural networks and deep learning for complex large-scale problems.
III. AI AND ITS RELEVANCE TO NTN CHALLENGES: Neural networks use layered neurons and weighted connections to transform inputs through intermediate representations into final outputs.

D. Major Learning Paradigms

The paper organizes learning into supervised, unsupervised, and reinforcement paradigms, then surveys representative ML, neural-network, generative, and sequence-model architectures. These approaches differ in supervision, representation learning, temporal modeling, dimensionality reduction, clustering, and data generation.

D. Major Learning Paradigms: Learning approaches are classified as supervised, unsupervised, or reinforcement learning according to training procedure and output-label availability.
D. Major Learning Paradigms: Supervised learning maps input features to known outputs, using iterative comparison between predictions and labels for tasks such as classification.
D. Major Learning Paradigms: Supervised ML includes linear and logistic regression and decision trees, while supervised deep learning includes perceptrons and fully connected neural networks.
D. Major Learning Paradigms: CNNs discover spatial features from multidimensional inputs through convolution and pooling, exploiting neighboring-cell correlation while reducing redundant features.
D. Major Learning Paradigms: RNNs capture temporal correlations through recurrent connections, while GRUs, LSTMs, and reservoir-computing variants address temporal or training-complexity considerations.Transformers provide encoder-decoder sequence-to-sequence architectures with embedding and positional encoding layers.
D. Major Learning Paradigms: Unsupervised learning discovers patterns in unlabeled data, using methods such as PCA, probabilistic graphs, K-means clustering, autoencoders, and generative models.Autoencoders encode and decode high-dimensional inputs to reconstruct them through compressed hidden representations.
D. Major Learning Paradigms: Self-organizing maps and GANs support competitive representation learning and data generation, while generative diffusion models iteratively denoise samples toward real-sample estimates.

3) Reinforcement Learning (RL):

Reinforcement learning models sequential decision-making through actions and feedback, with environments commonly represented as Markov Decision Processes. The survey distinguishes model-based, model-free, and deep extensions for complex NTN learning settings.

3) Reinforcement Learning (RL):: RL agents learn policies by taking actions in an environment and receiving rewards or penalties.The policy is the set of actions learned from experience.
3) Reinforcement Learning (RL):: Model-based RL assumes known state-transition probabilities, whereas model-free RL learns them iteratively.Dynamic Programming is identified as a prominent model-based method, while Q-learning and SARSA represent model-free approaches.
3) Reinforcement Learning (RL):: Deep reinforcement learning uses neural networks to estimate value functions or policies when RL state or action spaces are large.DQN, DDQN, Dueling DQN, Distributional DQN, and DRQN are described as representative architectures.
3) Reinforcement Learning (RL):: Distributed learning is relevant to satellite-based 6G because networks may combine large amounts of data from multiple operators.The survey highlights privacy and efficiency as inherent challenges of distributed approaches.
3) Reinforcement Learning (RL):: Federated learning shares locally trained models with a central server, while decentralized learning shares local models among neighboring nodes without exchanging actual data.Decentralized learning also avoids the centralized server required by federated learning.

1) Complex Task Automation:

AI is presented as a way to automate complex NTN operations that are difficult to optimize manually and can become computationally intractable. Its data-driven and adaptive models can support tractable, real-time decisions while reducing communication overhead.

1) Complex Task Automation:: NTN communication operations include resource allocation, channel estimation, modulation, coding, and satellite management control, making manual optimization often infeasible.The paper links this difficulty to the heightened complexity of satellite communication networks.
1) Complex Task Automation:: Multifaceted satellite architectures introduce many parameters, turning resource management into non-convex optimization problems that may require suboptimal or heuristic solutions.Deep learning can approximate complicated functions with neural networks to make such resource-management problems tractable.
1) Complex Task Automation:: AI models can adapt to time-varying NTN conditions by learning from experience and dynamically allocating resources, optimizing parameters, and detecting faults.The paper identifies reinforcement learning and predictive modeling as examples supporting online adaptation.
1) Complex Task Automation:: Data-driven AI reduces high-dimensional optimization difficulty through feature learning and can provide online decisions for latency-sensitive tasks such as scheduling and handover.NTN management decisions often need implementation within milliseconds to tens of milliseconds.
1) Complex Task Automation:: Existing CSI feedback can support reinforcement-learning models without modifying information segments sent from users to base stations.This reuse is especially relevant where NTN spectrum is scarce and expensive.

V. AI-NTN: CURRENT RESEARCH THRUSTS

The survey organizes AI-NTN research around challenges spanning communication layers, emphasizing channel estimation and Doppler shift in dynamic satellite environments. It presents machine learning as a practical alternative where conventional modeling is costly, delayed, or assumption-limited.

V. AI-NTN: CURRENT RESEARCH THRUSTS: AI-NTN research is categorized by communication-layer challenges, with lower-layer impairments receiving dedicated discussion.The survey focuses on satellite-based NTN because of satellites’ coverage, predictable trajectories, and scalability.
1) Channel Estimation:: Timely CSI acquisition is difficult in NTN because propagation delays and rapidly changing propagation environments hinder channel estimation.Channel estimation supports network planning and interference management.
1) Channel Estimation:: ML-based channel prediction can use distance, delay, received power, AoA, AoD, elevation angle, RMS delay spread, and frequency as inputs, with CSI as output labels.The survey cites ANN and LSTM-based approaches for channel and CSI prediction.
2) Doppler Shift Estimation:: Doppler shift arises from relative transmitter-receiver motion and changes the received signal frequency.The survey defines relative velocity and the angle between transceiver direction and signal propagation direction as equation variables.
2) Doppler Shift Estimation:: 48 kHz at a 2 GHz center frequency illustrates the significant frequency offset possible for LEO satellites.These offsets can cause UEs to tune away from assigned carriers and may create inter-carrier interference.
2) Doppler Shift Estimation:: Traditional Doppler estimators face orbital-mechanics complexity, simplifying assumptions, temporal variation, and ephemeris-related communication overhead.The survey presents ML-based algorithms as potential practical alternatives for characterizing Doppler effects.
2) Doppler Shift Estimation:: Channel state information can contain information about Doppler shift, enabling ML models trained from channel observations and ephemeris-derived ground truth.This connects CSI-based learning with Doppler estimation in mobile communication channels.

3) Security - Physical Layer Authentication:

The survey discusses physical-layer authentication and AI-based threat detection alongside beam-hopping optimization in satellite NTN. These topics address security vulnerabilities and the computational difficulty of managing many dynamic beams.

3) Security - Physical Layer Authentication:: Satellite-integrated interfaces enable spoofing and replay attacks, while large dynamic NTN topologies can make existing security assumptions and protocol overhead problematic.The text describes spoofing as satellite impersonation and replay as retransmission of intercepted messages.
3) Security - Physical Layer Authentication:: Physical-layer authentication verifies transmitter identity by exploiting distinctive channel characteristics between users and legitimate transmitters.The described encoding seeks to maximize mutual information between legitimate and wiretap channels.
3) Security - Physical Layer Authentication:: CNNs, autoencoders, and SVMs have been explored to extract authentication features from channel data, received signal power, and Doppler shift.These approaches are reported for legitimate-satellite authentication and radio fingerprinting.
3) Security - Physical Layer Authentication:: Threshold-based satellite threat detectors can produce false positives and miss anomalies with temporal correlations.The survey identifies deep learning as a response to these detection limitations.
3) Security - Physical Layer Authentication:: Beam hopping selects which beams to activate and for how long while optimizing metrics such as throughput, delay, and fairness under power and spectrum constraints.The problem is formulated as an optimization task whose objectives reflect network performance.
3) Security - Physical Layer Authentication:: Beam-hopping search complexity grows exponentially with beam count, and modern satellites may contain hundreds to thousands of beams.This scaling increases computation time for finding optimal patterns.
3) Security - Physical Layer Authentication:: Supervised learning and multi-agent deep reinforcement learning are used to learn beam-hopping patterns or dynamically allocate power and bandwidth.The cited approaches use labeled optimization outputs or DDQN-based cooperative agents.

2) Spectrum Sharing:

Integrated terrestrial–non-terrestrial networks reuse spectrum across hierarchical satellite and terrestrial systems, improving spectral efficiency but intensifying interference and allocation challenges. AI-based learning is explored alongside optimization to manage these difficult resource-sharing problems.

Spectrum Sharing: Shared satellite and terrestrial use of S and Ka bands can improve spectral efficiency and user QoE, but requires low-interference strategies.The integrated setting differs from traditional deployments where satellite and terrestrial networks generally occupy separate frequency bands.
Spectrum Sharing: Spectrum sensing detects target-band occupancy, but conventional methods trade low-SNR performance against computational complexity.Energy Detection is simple but performs poorly at low SNR, whereas CycloStationary and Eigen Value-based Detection offer better performance with greater complexity.
Spectrum Sharing: Learning approaches use historical occupancy, spatial-temporal correlations, and frequency-assignment data to support satellite spectrum sharing.Reported methods include CNN-BiLSTM, CNN-LSTM, and Q-Learning-based approaches for occupancy prediction, sensing, and assignment.
Spectrum Sharing: Joint power-and-spectrum allocation is necessary because spectrum non-orthogonality creates co-channel interference while power remains scarce and affects energy efficiency.Increasing power can suppress interference but cannot continue indefinitely without reducing energy efficiency.
Spectrum Sharing: The resulting optimization is often nonlinear, non-convex, and mixed-integer, motivating machine-learning methods that reduce conventional computational burdens.Deep learning has been combined with conventional optimization, while model-free DRL and Q-learning have been applied to satellite power and capacity allocation.

4) Network Slicing:

Satellite-terrestrial integration introduces resource-management problems spanning slicing, handover, multiple access, and power allocation. The surveyed work applies optimization, reinforcement learning, and deep learning to adapt these functions to changing service, mobility, and channel conditions.

Network Slicing: Network slicing partitions shared physical infrastructure into service-specific slices whose radio resources respond to user demand and traffic changes.Slicing can switch users between slices with different resource allocations while maintaining minimum service levels.
Network Slicing: Slicing optimization combines performance measures and costs under minimum-service constraints, while heuristic, reinforcement-learning, and neural approaches support practical implementations.Reported work covers slice creation, user association, scheduling, virtualized resource allocation, and FCNN-based approximation of nonlinear optimization.
Handover Optimization: LEO satellites remain visible to a ground user for only several minutes, requiring repeated handovers that challenge terrestrial attachment and load-balancing procedures.An LEO satellite travels at approximately 7.8 km/s and orbits Earth typically within two hours.
Handover Optimization: Greedy handover criteria such as maximum service time, maximum signal quality, and minimum network load provide heuristic solutions, while graph matching and reinforcement learning model broader objectives.RL formulations can treat handover criteria as states and user equipments as agents selecting suitable LEO satellites.
Multiple Access: NOMA and RSMA are investigated to improve spectral efficiency beyond conventional orthogonal access, but NOMA introduces SIC complexity and non-convex power allocation.NOMA shares time-frequency resources through power differentiation, whereas RSMA partitions messages into common and private components.

D. Upper Layer Aspects

Upper-layer satellite-terrestrial functions must coordinate computation, routing, and traffic adaptation under stringent delay, energy, and dynamism constraints. The survey identifies reinforcement learning for optimization and supervised learning for estimation as recurring patterns across these challenges.

Task Offloading: Task offloading must meet low-latency delay constraints while minimizing satellite energy consumption in integrated networks.The problem is formulated as an optimization task and has been addressed using hypergraph matching, game theory, stochastic methods, and efficient algorithms.
Task Offloading: Conventional offloading optimization can incur large network-state overhead and require many iterations, motivating decentralized and DRL-based alternatives.Reported approaches use channel state, dynamic satellite queues, DQN, DDQN, and DDPG, including consideration of security issues.
Network Routing: Satellite-terrestrial routing is difficult because hierarchical topology, link status, traffic, and channel conditions change dynamically, limiting direct use of shortest-path algorithms.Edge weights may represent delay, jitter, throughput, or packet loss, but simple Dijkstra routing does not directly meet integrated-network requirements.
Network Routing: Routing research combines adapted ATM and OSPF methods with static-dynamic schemes, ant-colony optimization, Kalman filtering, and coordinate-graph models.These approaches address dynamic satellite environments and multiple network constraints through different optimization formulations.
Traffic Prediction: Traffic prediction is especially important in NTNs because topology is highly dynamic and user requirements are diverse; deep learning captures spatial and temporal correlations.The surveyed methods include RBF networks, LSTM with attention, and GRU architectures designed to address recurrent-network gradient issues.
Key Takeaways: Reinforcement learning is mainly used for NTN optimization, whereas supervised learning is used for estimation tasks such as Doppler shift, channel state, and spectrum sensing.Unsupervised learning is described as less extensively covered because applying it in real networks is ambiguous and difficult.

VI. AI-NTN INTEGRATION: CURRENT STATUS

Current AI–NTN integration efforts combine satellite ML testbeds, O-RAN’s AI-control interfaces, and software-defined 5G-NTN implementations. The field is progressing toward practical deployment, but O-RAN and NTN standardization remain under development.

Current ML Testbeds: Existing satellite ML testbeds include MIRSAT for network-slicing experiments on NGSO constellations and ESA projects such as MLSAT and SATAI.These efforts provide experimentation platforms and investigate AI applicability to satellite networks.
O-RAN-Based RIC: O-RAN provides an open interface for exchanging RAN KPIs and control information with AI controllers, enabling closed-loop control and potential NTN integration.The framework is presented as a way to address the reconfiguration limits of traditional monolithic cellular networks.
O-RAN-Based RIC: Regenerative-payload NTN gNBs can place RU, DU, and CU across space or ground in three deployment configurations.Non-real-time functions are expected on the ground under power, onboard-capability, and mobility constraints, while near-real-time RIC placement depends on DU proximity.
SDR-Based Implementations: OAI and its FlexRIC support provide SDR-based, 3GPP-compliant experimentation with service-oriented controllers and NTN adaptations.OAI has been used for in-lab validation and over-the-satellite testing of 5G-NTN prototypes.
SDR-Based Implementations: Current implementations are mostly demonstrative, but NTN adaptations of OAI protocol stacks create a path for future AI deployment through xApps.The survey frames these software-defined efforts as experimental progress toward integrated satellite-terrestrial networks.
Key Takeaways: Satellite ML testbeds exist, while O-RAN is envisioned to support AI-enabled 6G NTN; both O-RAN and NTN standardization are still developing.The survey also notes ongoing incorporation of NTN adaptations into SDR-based 5G protocol stacks such as OAI.

VII. AI-NTN INTEGRATION: CHALLENGES

AI-enabled NTN integration is constrained by limited onboard resources, long propagation delays, security exposure, difficult scaling, scarce training data, and complex model tuning. These constraints are intensified by dynamic satellite-terrestrial network conditions and must be addressed for practical 6G deployment.

Resource constraints: Limited power, spectrum, and computational resources constrain onboard AI performance and large-scale NTN deployment.More capable AI hardware also increases power consumption, physical space, and maintenance cost.
Information and feedback: Long propagation delays hamper online reinforcement learning because timely environmental feedback is needed for rapidly changing network control.Slice and user resource allocation may require feedback on 1–10 ms and below 1 ms timescales, respectively.
Security: AI integration introduces attack surfaces including adversarial manipulation, data poisoning, model evasion, and denial-of-service attacks that can reduce network reliability.Attackers may consume excessive resources through extreme requirements in an accessed network slice, causing congestion.
Scalability: Large-scale satellite-terrestrial networks create high-dimensional state and action spaces, while multi-agent reinforcement learning state spaces grow exponentially with agent count.Although deep reinforcement learning can reduce state-space size, further work is needed for large-scale real-network deployment.
Data constraints: Quality training data can be costly, inefficient, or unavailable because of spectrum and intermittent-connectivity constraints, while non-terrestrial data distributions differ from terrestrial ones.These conditions can hamper training and degrade performance in real networks.
Model configuration: Neural-network hyperparameter tuning is uncertain and time-consuming, with NTN complexity making the relevant parameter space more difficult to configure.Training typically requires repeated empirical trials over parameters such as layers, activation functions, neuron counts, and learning rate.

J. Lack of Generalization

The survey reviews learning approaches and practical implementation requirements for AI-enabled satellite NTN integration in 6G. It emphasizes supervised and reinforcement learning, potential distributed approaches, RIC integration, and resource-aware deployment under severe operational constraints.

Existing Learning Approaches: Supervised learning suits estimation problems with well-labeled data, whereas reinforcement learning suits unlabeled NTN control problems with reward functions or network feedback.Examples include channel and Doppler-shift estimation for supervised learning, and resource allocation, beam hopping, and routing for reinforcement learning.
Existing Learning Approaches: Deep neural networks attract interest for satellite NTN problems because highly dynamic conditions and many variables can exceed the capabilities of traditional machine-learning approaches.Their feature-extraction capabilities support application to intricate network-related challenges.
Potential Learning approaches: Unsupervised learning can derive distributions of network parameters that may be inaccessible in real networks, while federated learning may reduce computing requirements.Highly dynamic, time-varying NTN behavior remains difficult for unsupervised approaches to capture.
Enabling O-RAN-based RIC: Software-defined NTN prototypes are being adapted to OAI 4G and 5G stacks, but integrating a RAN Intelligent Controller is described as crucial for deploying AI algorithms.The paper links RIC integration with providing an interface for AI in real NTN implementations.
Practical Implications: Practical AI-enabled NTN deployment must contend with onboard-computation cost, environmental extremes, propagation delay, power, bandwidth, security, and physical-space limits.The paper points to miniaturization, secure design, energy efficiency, and careful bandwidth use as relevant enabling directions.

1) Interrelated Issues:

NTN challenges are interconnected, so practical AI-enabled solutions must jointly account for network dynamics, online computation, distributed learning, feedback overhead, onboard constraints, energy, and security.

1) Interrelated Issues:: NTN issues are interdependent; for example, incorporating resource allocation into handover decisions can improve network performance.Network load changes after user attachment, linking handover and resource allocation decisions.
2) Recurrent Learning Architectures:: Recurrent architectures can better capture temporal dependencies in beam-hopping, resource allocation, and network slicing than feed-forward networks.The paper identifies RNNs and related architectures as suitable for time-varying NTN problems, with ESNs suggested for dynamic spectrum access and sharing.
3) Online Implementation:: Online NTN control is limited by the computational complexity of current algorithms, making complex deep feed-forward networks impractical for real-time decisions.The paper proposes investigating ESNs, ELMs, or combinations with feed-forward networks for more viable online implementation.
4) Distributed Learning Models:: Distributed learning across multiple computing nodes can improve scalability, computation speed, and efficiency in integrated satellite-terrestrial networks.Candidate approaches include data parallelism, model parallelism, ensemble learning, and federated learning.
5) Control Feedback Design:: NTN feedback mechanisms must reduce overhead while supporting network optimization and efficient resource allocation.Combined feedback can serve multiple AI approaches and improve the effectiveness of NTN learning algorithms.
6) Development in Miniaturization:: Miniaturization can increase satellite onboard capability by enabling more powerful processors and larger memory within limited space.This depends on advances in materials, integrated circuits, systems-on-chip, and MEMS design.
7) Energy System Design:: Satellite launches and maintenance consume substantial power, making lightweight components, power management, conversion, propulsion, and storage important energy-design considerations.Energy efficiency is identified as a critical criterion for NTN platforms.
8) Secured System Design:: Secure NTN design requires intrusion detection and prevention tailored to attacks that can degrade performance and compromise data integrity and confidentiality.Continuous monitoring of relevant network parameters enables anomaly detection and prompt mitigation.