Regenerative Satellite Payloads

• Regenerative payloads are onboard processing architectures that demodulate, decode, and digitally regenerate signals, enabling adaptive satellite communications.
• They support advanced functions like digital beamforming, multi-spot multiplexing, and QoS provisioning for broadband direct-to-device and 6G non-terrestrial scenarios.
• Modern designs integrate SDN, virtualization, and joint resource optimization to enhance latency, user admission, and overall system efficiency in high-throughput networks.

A regenerative payload in satellite systems denotes an on-board architecture that performs demodulation, decoding, regeneration, and flexible digital processing—contrasting with “bent-pipe” payloads, which simply amplify and forward signals. Regenerative payloads enable in-orbit adaptation, intelligent resource management, and programmability, supporting broadband direct-to-device (D2D) and non-terrestrial 6G scenarios, multi-spot-beam spatial multiplexing, on-orbit Quality of Service (QoS) provisioning, caching and multicasting, and dynamic network function placement. Recent designs leverage distributed, software-defined, and scalable structures, including modem banks, on-board radio access network (RAN) functions (e.g., gNB, DU), and advanced joint optimization of bandwidth, power, and beam coverages.

1. Fundamental Architecture and On-Board Digital Signal Chain

Regenerative payloads incorporate on-board digital processing to terminate, demodulate, decode, switch, and re-encode user traffic before retransmission. Core architectural elements include:

• RF and Analog Front-End: Each uplink and downlink beam is routed through dedicated front-ends. Signals are downconverted, digitized via ADCs, and pre-processed.
• Digital Regenerative Processor: Implements demodulation, channel decoding, packetization, digital beamforming, spatial multiplexing, QoS queuing, caching, and per-flow packet switching.
• Programmable Modem Banks: Multiple banks, each terminating (receiving, regenerating) one or more beams, laid out in toroidal (e.g., 4×4) mesh topologies. Each bank integrates DSP, local memory/buffers, and four gigabit interconnects to neighbors (Yahia et al., 2024).
• SDN/Network Function Virtualization: Embedded general-purpose processors execute a Linux-based software stack running SDN agents for routing, load balancing, and QoS enforcement, enabling full software-defined and reconfigurable payload operation (Yahia et al., 2024, Kong et al., 9 Sep 2025).

Comparative Table: Bent-Pipe vs. Regenerative Payload

Feature	Bent-Pipe Payload	Regenerative Payload
Demodulation/Decoding	Ground	On-board
Routing/Switching	Fixed-path/Frequency routing	Software-defined, per-packet routing
Beamforming/Spatial Multiplex	Static/Analog	Fully digital, adaptive
Caching/Content Delivery	Not available	On-board policy- and cache-driven
Flexibility	Rigid	Dynamically reconfigurable

As established in (Yahia et al., 2024, Bhandari et al., 2023).

2. Forms of Regenerative Payloads in Satellite Access Networks

Modern regenerative payloads may instantiate different functional splits, prominently:

• On-board gNB (“Split 0”): The full set of radio access (CU and DU) functions, i.e., PHY, MAC, RLC, PDCP, and RRC, is executed in orbit.
• On-board gNB-DU (“Split 2”): Only PHY, MAC, RLC reside onboard. The satellite communicates with a ground-based CU (PDCP, RRC) via the 3GPP F1 interface (Kong et al., 9 Sep 2025).

On-board gNB enables lowest end-to-end latency but highest on-board compute (OPEX), while gNB-DU reduces OPEX via offload but incurs feeder-link F1 delay (≈80–100 ms).

Payload Type	RTT (ms)	CPU Utilization (%)
On-board gNB	61.4	70
On-board gNB-DU	145.1	50

Measured in OAI-based prototype in (Kong et al., 9 Sep 2025).

This tradeoff is fundamental: flexible payload architectures may support dynamic per-user or per-flow function placement to optimize either cost or QoS in real time (Kong et al., 9 Sep 2025).

3. Joint Resource Management and Optimization Formulations

Regenerative payloads enable joint optimization over multiple resource and functional domains, including routing, bandwidth, power, user grouping, and function placement:

• Flow Routing & Load Balancing: The inter-bank payload is modeled as a directed graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$, where nodes are modem banks and links are gigabit-class interconnects. Traffic is managed via a max–min residual capacity problem, maximizing the minimal link slack and thus minimizing hot spots and delay (Yahia et al., 2024).
• QoS-aware Function Placement: In FlexSAN, P₁ (OPEX minimization) and P₂ (service maximization) are solved as mixed-integer nonlinear programs (MINLPs) or via the TAGO heuristic. Constraints include per-UE delay $t_u \leq T_u^{max}$, rate $R_u^s \geq R_u^{min}$, per-user resource exclusivity, bandwidth, and total sat. processing constraints. The objective alternates between total OPEX (GOPS) and number of admitted UEs (Kong et al., 9 Sep 2025).
• Multidimensional Radio Resource Optimization: In FLARE-LEO, radio resources $w$, $b$, $r$, $K$ (precoding weights, bandwidth per active user group, beam radii, user group cardinality) are jointly optimized to minimize the worst-case delivery latency under constraints on per-beam power, bandwidth, and coverage. K-means clustering is used for spot-beam shaping, and SCA-based iterative algorithms for bandwidth/precoding allocation (Bhandari et al., 2023).
• Handover Enhancements: During LEO handover, joint transmission architectures—centralized or distributed—use deep learning-based CSI prediction and collaborative power control to maintain rate and minimize delivery delay (Bhandari et al., 2023).

4. Algorithmic Techniques and Implementation

Scalable and real-time resource orchestration for regenerative payloads relies on computationally efficient heuristics:

• TAGO (Two-stage Adaptive Greedy Orchestration): For FlexSAN, congestion state is characterized by a scalar score $\sigma$ combining processing and bandwidth utilization. TAGO invokes cost-efficient (CEO) or service-maximizing (SMO) phases depending on load, prioritizing users via composite delay-risk and CPU-cost metrics, and performing bandwidth compression or function swapping as needed. Overall time complexity is $O(N \log N)$. Achieves near-optimal admission within 5.5% of Gurobi at >100× speedup (Kong et al., 9 Sep 2025).
• Successive Convex Approximation (SCA): For radio resource allocation, nonconvex constraints (e.g., SINR or rate expressions) are recast via successive convex inner approximations, typically converging within 5 iterations (Bhandari et al., 2023).
• K-means Beam Clustering: Spatial spot beam coverages are adapted to traffic geography by allocating UEs via K-means clustering, then setting beam radii to span assigned users (Bhandari et al., 2023).
• CNN-based CSI Prediction: During LEO handover, 2D CNNs predict future channel states based on historical measurements, using pretraining/transfer learning between gateway and on-orbit platforms (Bhandari et al., 2023).

5. Performance Outcomes and Comparative Evaluations

• Latency and Admission: FlexSAN achieves a 36.1% average increase in user admission and 15% OPEX reduction against static payload schemes. Under high-stress scenarios (strict delay and heavy load), dynamic functional splitting maintained 63.5% admission vs. 37.8% for static schemes. Near-instant reconfiguration is possible (<100 ms orchestration) (Kong et al., 9 Sep 2025).
• Delay and Loss: Toroidal modem-bank architectures (16-bank, 10 Gbit/s links) saw sub-1 ms delays and negligible packet loss at offered loads up to 90 k packets/s, whereas even 4×-scaled monolithic banks exceeded 5 ms delay and >10% loss at such loads (Yahia et al., 2024).
• Resource Adaptation: Adaptive beamforming using K-means delivered effective mean/min user rates ≥1.22× those of fixed coverage. SCA precoding achieved 10–20% throughput improvements vs. zero forcing. Caching brought up to 50% delivery time reduction depending on library fraction cached (Bhandari et al., 2023).
• Handover Enhancements: Centralized collaborative handover provided ∼1.5× throughput boost over non-HO, and deep learning predictions enabled robust operation with MSE under $3.8 \times 10^{-4}$ on correlated channels (Bhandari et al., 2023).

6. Practical Considerations and Prospective Research Directions

Key implications and outstanding challenges for regenerative payloads include:

• Power, Mass, and Fault Tolerance: Distributed banks and software-defined switching architectures avoid large, centralized crossbar switches, reducing mass and power. Toroidal topologies enable alternate-path rerouting, supporting graceful degradation under hardware failure (Yahia et al., 2024).
• Network Function Programmability: Dynamic per-user functional split, per-flow resource allocation, and the ability to instantiate/shift RAN components on-orbit are critical for future non-terrestrial networks (Kong et al., 9 Sep 2025).
• Integration with Machine Learning: Predictive resource orchestration (e.g., for traffic or outage forecasting) and real-time channel state estimation are recognized as essential, but practical, inter-satellite load balancing and integration with ground-side orchestration remain active research areas.
• Scalability: Scaling throughput is more readily achieved via additional banks and higher link rates, rather than unfeasible monolithic switches. The architecture is suited to large LEO constellations with direct-to-cell services (Yahia et al., 2024).
• Limitations: Current solutions focus on single-satellite coverage; extension to multi-satellite, dynamic inter-satellite networking is identified as a priority for future work (Kong et al., 9 Sep 2025, Bhandari et al., 2023).

The emergence of regenerative payloads is thus central to the performance, flexibility, and economic viability of next-generation high-throughput satellite and non-terrestrial 6G networks, enabling architecture-level programmability, real-time adaptation, and efficient broadband service delivery under demanding spatiotemporal traffic scenarios (Yahia et al., 2024, Kong et al., 9 Sep 2025, Bhandari et al., 2023).