Satellite Channel Emulator Overview

• A satellite channel emulator is a system that replicates satellite link propagation, impairment, and network effects across multiple abstraction levels.
• It integrates advanced methods like differentiable stochastic channel layers, I/Q-level SDR digital twinning, and trace-driven network emulation for comprehensive system analysis.
• These emulators support applications from neural source–channel coding to 5G NTN testing and digital twin network analysis, while highlighting challenges in reproducing real-world variability.

A satellite channel emulator is a system that reproduces, in a controlled and repeatable form, the propagation, impairment, and network effects of satellite links for analysis, training, validation, or protocol experimentation. In recent literature, the term spans several abstraction levels: differentiable stochastic channel layers inside neural joint source–channel coding pipelines; I/Q-level digital twins that sit between software-defined radios and virtual radios; SDR-based end-to-end non-terrestrial network platforms that exercise a real satellite path; packet-level digital twins that map geometry, interference, and routing into time-varying delay and capacity; and laboratory optical benches that replay attenuation, beam wander, and turbulence for free-space optical payloads (Kondrateva et al., 1 Aug 2025, Gurses et al., 6 Apr 2026, Hou et al., 25 Sep 2025, Gao et al., 2024, Medlock et al., 26 Feb 2026).

1. Scope and abstraction levels

The modern literature does not treat a satellite channel emulator as a single device class. Instead, it is a family of instruments and software frameworks whose common function is to expose a communication system to satellite-specific channel behavior. At the lowest abstraction, the emulator acts on complex baseband samples or optical fields. At higher abstraction, it acts on packet streams, routing state, or even multicast erasure processes. This breadth is not merely terminological: it reflects the fact that different research questions require different notions of “the channel,” ranging from symbol-level fading to end-to-end path dynamics (Kondrateva et al., 1 Aug 2025, Ghanem et al., 2017).

Abstraction level	Representative systems	Emulated quantities
Differentiable/statistical baseband	DJSCC-SAT, ADJSCC-SAT	Fading, AWGN, link-budget SNR, environment/state dependence
I/Q-level SDR digital twin	ACHEM, SDR-based 5G NTN platform	Delay, frequency offset, CIR, antenna gain, timed sample delivery
Packet/network level	Plotinus, NetSatBench, OpenSN, Hypatia trace replay	Connectivity, delay, capacity, queueing, routing, handover
Optical bench	SPOQC optical emulator	Dynamic loss, beam wander, wavefront distortion, scintillation

A further abstraction appears in multicast coding, where the “channel” is virtualized as a single worst-case receiver. In the MaxPe scheme, the virtual packet-erasure process is the per-slot maximum erasure among receivers, $P_e^{(v)}(t)=\max_k P_e(h_k(t))$; in MaxCT, the virtual channel is the channel of the receiver with maximum completion time. This is a network-layer form of channel emulation, because the multicast group is replaced by one equivalent receiver that drives code adaptation (Ghanem et al., 2017).

2. Statistical and differentiable RF emulation

A prominent RF formulation is the differentiable channel layer used for deep joint source–channel coding in small-satellite downlinks. In DJSCC-SAT, the encoder maps an input image $x\in\mathbb{R}^n$ to $\tilde z\in\mathbb{C}^k$, power normalization enforces an average transmit-power constraint, the channel layer applies

$\hat z = z h + n,$

and the decoder reconstructs $\hat x$. The emulator is explicitly implemented as a non-trainable, stochastic, differentiable layer between encoder and decoder, so that end-to-end training can treat channel variation as part of the data distribution (Kondrateva et al., 1 Aug 2025).

The fading law is a multi-state narrowband land-mobile satellite model based on the Fontán framework. Three shadowing states are used—LOS, shadow, and deep shadow—combined with five environment classes: open, suburban, intermediate tree shadow, heavy tree shadow, and urban. In the reported experiments, the states are treated independently rather than through an explicit Markov evolution, so the emulator is memoryless at the symbol-block level even though its statistics are derived from a model that originally includes a three-state Markov chain (Kondrateva et al., 1 Aug 2025).

Within each state, the complex gain follows a Loo distribution: a log-normal direct component plus a Rayleigh multipath component. The emulator uses state-, environment-, and elevation-dependent parameters $(\alpha,\psi,\Omega)$, with $|h_{\text{LOS}}|\sim\text{LogNormal}(\alpha,\psi)$ and $h_{\text{NLOS}}\sim\mathcal{CN}(0,\Omega)$. Noise is generated as complex AWGN with variance derived from a link-budget SNR,

$\mathrm{SNR} = P_t + G_t + G_r - L - N,$

with path loss from a Friis formulation and slant range computed from orbit height, Earth radius, and elevation angle. In the 2025 framework, the representative link parameters are orbit height $750$ km, carrier frequency $x\in\mathbb{R}^n$0 MHz, transmit power $x\in\mathbb{R}^n$1 W, satellite antenna gain $x\in\mathbb{R}^n$2 dBi, ground antenna gain $x\in\mathbb{R}^n$3 dBi, bandwidth $x\in\mathbb{R}^n$4 kHz, and noise figure $x\in\mathbb{R}^n$5 dB (Kondrateva et al., 1 Aug 2025).

The adaptable extension, ADJSCC-SAT, uses channel-aware attention modules to avoid training one neural network per channel condition. Each attention block concatenates global-average-pooled feature descriptors with the channel parameter vector

$x\in\mathbb{R}^n$6

passes the result through a two-layer MLP, and applies channel-wise scaling to the feature maps. The additional attention parameters are only about $x\in\mathbb{R}^n$7 of the total parameter count. Empirically, the adaptable model approaches the performance of multiple specialized networks, shows enhanced robustness to SNR and state mismatch, and reduces model storage requirements, which is particularly relevant under onboard constraints (Kondrateva et al., 1 Aug 2025).

An earlier adaptable JSCC formulation treated the emulator in the same architectural role—the middle, non-trainable layer of an autoencoder—and similarly emphasized a complex-baseband Loo-plus-AWGN model with environment-dependent parameters and a potential Markov shadowing interpretation (Kondrateva et al., 2024).

3. I/Q-level and SDR-integrated realization

At I/Q level, ACHEM provides a software-based end-to-end channel emulator for UHD-based SDR systems. Its architecture consists of V-USRP, a virtual USRP implemented inside UHD, and CHEM, the channel emulator that exchanges signal frames with V-USRP over UDP. Any UHD-based application can therefore run unmodified while the physical radio is replaced by a digital channel path (Gurses et al., 6 Apr 2026).

The baseband model used by ACHEM is

$x\in\mathbb{R}^n$8

which includes path loss, time-varying CIR taps, antenna radiation patterns, propagation delay in samples, frequency offset, and additive noise. The framework supports multiple transmitters and receivers, MIMO communications, multiple frequencies, heterogeneous sampling rates, and 3D mobility with orientation. Validation was reported with GNU Radio, srsRAN 4G/5G, and OpenAirInterface; for single-link LTE and NR profiles, measured end-to-end CHEM processing latencies were on the order of $x\in\mathbb{R}^n$9–$\tilde z\in\mathbb{C}^k$0, and V-USRP latency remained below $\tilde z\in\mathbb{C}^k$1 ms for up to 16 concurrent sources (Gurses et al., 6 Apr 2026).

A distinct but related realization appears in the SDR-based 5G NTN platform built around Amarisoft protocol stacks and FlexSDR400 hardware. That platform is not a black-box channel emulator; it uses a real GEO transparent payload, AsiaSat-9, as the channel and therefore functions as an end-to-end NTN testbed whose SDR and stack flexibility emulate many roles of a channel emulator. The channel characteristics observed and compensated include long GEO propagation delay, a cumulative frequency offset of about $\tilde z\in\mathbb{C}^k$2–$\tilde z\in\mathbb{C}^k$3 kHz, and NTN-specific timing advance based on Release-17 signaling. In field trials, the measured PDSCH SNR was about $\tilde z\in\mathbb{C}^k$4 dB, downlink throughput about $\tilde z\in\mathbb{C}^k$5 Mbit/s over $\tilde z\in\mathbb{C}^k$6 MHz, BLER below $\tilde z\in\mathbb{C}^k$7 on average, and RTT about $\tilde z\in\mathbb{C}^k$8 ms on average, with minima around $\tilde z\in\mathbb{C}^k$9 ms and maxima around $\hat z = z h + n,$0 ms (Hou et al., 25 Sep 2025).

The GEO testbed demonstrates that emulating a satellite channel for NTN cannot be reduced to propagation delay alone. Protocol timing, timing advance,

$\hat z = z h + n,$1

PRACH search range, and HARQ behavior become part of the effective channel seen by the system (Hou et al., 25 Sep 2025).

4. Digital twins, path replay, and network-scale emulation

At packet and network level, the emulator becomes a digital twin of constellation geometry, interference, routing, and application behavior. Plotinus is representative of this class. It is a microservice-based digital twin built around NS-3 v3.31, with pluggable physical-layer, topology, path computation, and visualization services. Its physical-layer plugin computes Bessel-function-based beam gain $\hat z = z h + n,$2, channel coefficient

$\hat z = z h + n,$3

interference

$\hat z = z h + n,$4

and capacity

$\hat z = z h + n,$5

Those time-varying capacities are fed into NS-3, while TapBridge connects virtual nodes to real containers or hosts so that live traffic experiences the emulated network in real time (Gao et al., 2024).

NetSatBench and OpenSN push this network-emulation paradigm to large LEO constellations. NetSatBench implements satellites, gateways, and user terminals as Linux containers connected through a Layer-2 VXLAN overlay, stores state in Etcd, and updates link parameters through epoch files. Its physical-layer and routing models are decoupled into external plug-ins, while built-in support exists for IPv4/IPv6 routing, IS-IS, and ideal time-varying routing. A representative satellite-ground bitrate law is

$\hat z = z h + n,$6

with slant-range-dependent attenuation (Detti et al., 30 Apr 2026).

OpenSN likewise emulates satellites, ground stations, and users as containers, but emphasizes control-plane decoupling through Etcd, direct netlink-based link management, and eBPF/XDP links. It can emulate the five-shell Starlink constellation with a total of $\hat z = z h + n,$7 satellites, constructs mega-constellations $\hat z = z h + n,$8–$\hat z = z h + n,$9 faster than StarryNet, and updates link state $\hat x$0–$\hat x$1 faster than LeoEM. Its role as a channel emulator is packet-level: per-link delay, bandwidth, and connectivity are updated according to orbital geometry and handover policy, while real routing software runs inside each node (Lu et al., 4 Jul 2025).

A complementary architecture is trace-driven path emulation using Hypatia. There, Hypatia simulations record end-to-end path characteristics such as delay and bandwidth between two ground stations into Trace Files, which are then replayed in real time by a kernel-level qdisc, TheaterQ. The replayed quantities include one-way delay, available path capacity, queue capacity, packet loss, and a route identifier marking path changes. Reported correlation metrics between full simulation and emulation reach up to $\hat x$2 for TCP goodput, validating the trace-driven approach while also exposing synchronization issues and the lack of emulation-to-simulation feedback (Ottens et al., 30 Oct 2025).

The limitations of current constellation emulators are now empirically documented. A side-by-side evaluation of StarryNet, OpenSN, and Celestial against real Starlink measurements from WetLinks concluded that none of them reproduces realistic RTT and throughput variability, because their link/channel models remain too simplistic and constellation updates can lag behind real time (Cionca et al., 6 Apr 2026).

5. Optical, terahertz, and ray-tracing-derived emulators

Free-space optical emulation is represented by the SPOQC channel emulator for continuous-variable quantum key distribution. That laboratory apparatus replays a LEO satellite-to-ground optical downlink using a variable optical attenuator, a fine steering mirror, a Galilean beam expander, and a deformable mirror driven by Zernike-decomposed turbulence phase screens. It emulates dynamic path loss, beam diffraction, pointing loss, atmospheric scattering, beam wander, wavefront distortion, and scintillation for a $\hat x$3 km LEO pass. The underlying model includes a Hufnagel–Andrew–Phillips $\hat x$4 profile, atmospheric transmission

$\hat x$5

beam broadening, and beam-wander variance

$\hat x$6

The VOA updates at about $\hat x$7 Hz, while the FSM and deformable mirror operate up to $\hat x$8 kHz, matching the distinct time scales of pass loss and turbulence (Medlock et al., 26 Feb 2026).

At terahertz frequencies, emulator design is guided by a deterministic-plus-stochastic path model for ULEO satellite-to-ground links: $\hat x$9 with curved-Earth atmospheric absorption integrated over altitude and optional Weibull fading. The paper proposes direct satellite-to-ground, satellite–relay–ground, and satellite-to-high-altitude-base-station architectures; for the fading case it uses Weibull parameters $(\alpha,\psi,\Omega)$0 and $(\alpha,\psi,\Omega)$1. This formulation explicitly suggests an emulator block structure consisting of geometry/FSPL, atmospheric absorption, deterministic attenuation combination, fading, and AWGN (Zhang et al., 16 Sep 2025).

Ray-tracing-derived models provide another route to emulator parameterization. For urban LEO satellite-to-ground communication at $(\alpha,\psi,\Omega)$2 GHz, a shooting-and-bouncing-rays model with SGP4 orbital dynamics yields time-varying received power, RMS delay spread, and Doppler over the visibility window. A key result is that multipath Doppler shifts differ by only $(\alpha,\psi,\Omega)$3–$(\alpha,\psi,\Omega)$4 Hz while the absolute Doppler reaches about $(\alpha,\psi,\Omega)$5 kHz, so a common Doppler term per snapshot is often a reasonable emulator approximation (Ning et al., 6 Jan 2025). For suburban X-band links at $(\alpha,\psi,\Omega)$6 GHz, Wireless InSite simulations lead to an elevation-aware large-scale model that adds hardware loss, atmospheric attenuation, and antenna misalignment to coherent ray-traced power, while small-scale fading is fitted by shadowed Rician near the horizon and Rician or pure LOS above it. Reported RMS delay spreads are very small—90% of values below $(\alpha,\psi,\Omega)$7 ns for the $(\alpha,\psi,\Omega)$8 km pass and below $(\alpha,\psi,\Omega)$9 ns for the $|h_{\text{LOS}}|\sim\text{LogNormal}(\alpha,\psi)$0 km pass—so many such links are effectively flat fading unless bandwidth is very large (Khawaja et al., 19 Jul 2025).

6. Applications, validation, and open limitations

Satellite channel emulators are used to answer different classes of questions, and their validation metrics follow those questions. Neural source–channel systems evaluate reconstruction metrics such as MSE and PSNR under controlled fading states and SNR mismatch (Kondrateva et al., 1 Aug 2025). NTN SDR platforms evaluate downlink throughput, BLER, and RTT under realistic timing-advance and HARQ behavior (Hou et al., 25 Sep 2025). Digital twins such as Plotinus, OpenSN, and NetSatBench expose RTT, hop count, path changes, flow completion, and congestion-control behavior to real applications and routing daemons (Gao et al., 2024, Lu et al., 4 Jul 2025, Detti et al., 30 Apr 2026).

A network-centric GEO example is OpenSAND-based LTE backhaul emulation, where the satellite system is abstracted through capacity, delay, and return-link access mechanisms rather than detailed RF fading. The reported forward and return capacities are $|h_{\text{LOS}}|\sim\text{LogNormal}(\alpha,\psi)$1 Mbps and $|h_{\text{LOS}}|\sim\text{LogNormal}(\alpha,\psi)$2 Mbps, and CRA/RBDC settings are varied to study QoE for web browsing, file transfer, and VoIP. This suggests that, for transport and QoE studies, the decisive channel variables may be asymmetry, delay, and scheduler dynamics rather than waveform-level realism (Nicolas et al., 2021).

Emulator inputs can also be generated from traffic and beam models. A multibeam GEO traffic simulator combines population, aeronautical, and maritime datasets with realistic beam patterns, then builds a complex channel matrix

$|h_{\text{LOS}}|\sim\text{LogNormal}(\alpha,\psi)$3

whose entries are derived from beam gain, free-space path loss, receive antenna gain, and path-length-dependent phase. Such models are natural front ends for emulators concerned with precoding, beam hopping, and interference-aware scheduling (Al-Hraishawi et al., 2020).

Across the literature, limitations are consistent. The DJSCC-SAT emulator is narrowband and flat fading, omits Doppler, time correlation, and hardware impairments, and treats the original Fontán Markov structure only statistically (Kondrateva et al., 1 Aug 2025). SLASh, a LEO LISL simulator, assumes ideal, always-available LISLs with fixed bitrate and static satellite positions during a run, making it useful for topology studies but not for RF channel emulation (Romine et al., 27 Dec 2025). Trace-driven path replay inherits the lack of emulation-to-simulation feedback and is sensitive to start-time synchronization (Ottens et al., 30 Oct 2025). Empirical comparison against WetLinks shows that current open-source constellation emulators still fail to reproduce realistic Starlink-like channel variability because they largely reduce the channel to geometry-derived delay plus static loss (Cionca et al., 6 Apr 2026).

The general direction of the field is therefore convergent rather than contradictory. Statistical emulators are moving toward richer state dependence and side-information conditioning; SDR-based digital twins are moving toward full I/Q transparency with programmable mobility and antenna models; network emulators are decoupling orbital dynamics, routing, and effective link models through plug-ins and trace interfaces; and optical or high-frequency emulators are becoming increasingly physics-driven. A plausible implication is that future satellite channel emulators will be hybrid systems: ray-tracing- or measurement-derived propagation blocks feeding differentiable or real-time replay engines, coupled to protocol-aware digital twins that preserve the timing and control semantics of the target stack (Kondrateva et al., 1 Aug 2025, Gurses et al., 6 Apr 2026, Gao et al., 2024, Ottens et al., 30 Oct 2025).