©2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
HELENA for 5G NR LEO NTN Channel Estimation: A Comparative Evaluation
arXiv:2609.14735v1 [eess.SP] 13 Sep 2026
Miguel Camelo Botero, Nina Slamnik-Kriještorac, Johann Marquez-Barja University of Antwerp - imec, IDLab, Antwerp, Belgium
Abstract—Deep Learning (DL)-based channel estimation has shown high accuracy and low latency in terrestrial 5G NR, but Low Earth Orbit (LEO) Non-Terrestrial Networks (NTNs) introduce Doppler and synchronization impairments that may require NTN-specific architectures. We test whether HighEfficiency Learning-based channel Estimation using dual Neural Attention (HELENA), originally designed for terrestrial channels, remains effective after NTN retraining and suitable across highperformance and power-constrained inference platforms. Its unchanged architecture is trained on paired receiver-compensated (NTN-1) and residual-impaired (NTN-2) datasets and compared with eight terrestrial-origin models trained on the same NTN data and the NTN-specific MDELAN-SISO. HELENA achieves the lowest observed SNR-averaged NMSE among the DL estimators in both conditions, including 55.8–62.7% lower linear-scale NMSE than MDELAN-SISO. All DL models degrade in NTN2, demonstrating the challenge posed by residual Doppler and its associated impairments. On an RTX PRO 4500, HELENA achieves 0.0595 ms 99th-percentile (P99) inference latency, 88.1% below the 0.5 ms budget, with lower energy than its closest attention-based competitors. On a 10 W Jetson Orin NX, it retains a favorable accuracy–energy trade-off, but no model meets the P99 budget. Thus, HELENA needs no NTN-specific redesign for the evaluated task, while embedded tail latency remains an open challenge. Index Terms—Channel Estimation, Non-Terrestrial Networks, LEO Satellites, Deep Learning, Neural Network Acceleration.
I. I NTRODUCTION Fifth-generation New Radio (5G NR) Non-Terrestrial Networks (NTNs) use satellites or airborne platforms to extend cellular access to remote, maritime, and aeronautical areas. In Low Earth Orbit (LEO) access, orbital motion creates large Doppler shifts and time-varying propagation delays, while UEs at different positions within a satellite beam experience different Doppler and delay conditions [1]. These impairments complicate synchronization and channel estimation compared to terrestrial 5G NR. To mitigate these effects, the transmitter can use ephemeris and a beam-reference position to pre-compensate the Doppler component common to the beam [2]. A UE away from this reference retains a location-dependent frequency offset, further modified by its mobility. The receiver must first acquire timing and frequency and may then estimate this residual using cyclic-prefix redundancy and NR reference signals [3], [4]. Imperfect synchronization leaves phase evolution and inter-carrier interference in the grid presented to the channel estimator. Consequently, a practical NTN estimator must recover the effective post-processing channel rather than only
the underlying multipath channel commonly considered in terrestrial evaluations [5], [6]. Pilot-aided Least Squares (LS) estimation is inexpensive but noise-sensitive, whereas Linear Minimum Mean Square Error (LMMSE) improves robustness using second-order statistics at higher acquisition and computational cost [7], [8]. Learning-based terrestrial estimators include convolutional super-resolution and residual networks [5], [9]–[11], fully connected models [12], transformers [13], and model-driven receivers such as MDX [14]. More recently, Multi-Dilated Efficient Layer Aggregation Network (MDELAN) [15] extends MDX to NTN with a specific focus on channel estimation. High-Efficiency Learning-based channel Estimation using dual Neural Attention (HELENA) combines sparse pilot input with lightweight dual attention and was originally designed for terrestrial OFDM channel estimation [6]. It remains unclear whether its unchanged architecture retains its accuracy– efficiency advantages after NTN retraining, particularly under different residual-Doppler conditions and across highperformance and power-constrained embedded accelerators. To investigate this, we retrain HELENA without architectural modification and compare it with eight terrestrial DL estimators retrained from scratch and one NTN-specific estimator. Paired receiver-compensated (NTN-1) and residualimpaired (NTN-2) datasets isolate the impact of residual Doppler, while FP16 deployment on high-performance and embedded platforms evaluates latency and energy efficiency. In summary, the main contributions are: a) paired receivercompensated and residual-impaired datasets isolating receiver Doppler effects; b) a common comparison of HELENA with eight terrestrial DL estimators, one NTN-specific estimator, and practical pilot-based and statistical baselines, showing the lowest observed DL SNR-averaged NMSE in both NTN conditions; and c) an accuracy–complexity–latency–energy analysis using FP16 TensorRT on high-performance and powerconstrained embedded accelerators. The paper continues as follows: Sections II–IV present the system model, methodology, and dataset generation, respectively. Section V reports the results, Section VI discusses their implications, and Section VII concludes the paper. II. S YSTEM M ODEL AND P ROBLEM S TATEMENT We consider a downlink Single Input Single Output (SISO) 5G NR Orthogonal Frequency-Division Multiplexing (OFDM) link between a LEO satellite and a terrestrial User Equipment (UE). Let fd (θ) denote the signed satellite-induced Doppler
shift, in hertz, observed at a ground location with satellite elevation angle θ. In Fig. 1, the beam-center reference (BC) is located at the footprint center, while the UE may occupy another point in the beam. Their elevation angles θBC and θUE define the beam-common Doppler fd,common = fd (θBC ) and the satellite Doppler at the UE fd,sat = fd (θUE ), respectively. The transmitter is assumed to know fd,common from the satellite and beam-reference geometry and removes it before radiation. The residual carrier-frequency offset at the receiver input is then fd,pre = fd,sat − fd,common + fd,UE ,
(2)
pre where Xi,k and Yi,k
are the transmitted and received grid can symbols, respectively. Hi,k is the noise-free multipath response in the transmitter reference frame, excluding the residual Doppler and timing transformations represented by the exponential terms. Moreover, τ is the timing error relative to the OFDM demodulation window, and Zi,k ∼ CN (0, σ 2 ) is circularly symmetric complex Gaussian noise with total complex variance σ 2 = E[|Zi,k |2 ] per resource element. The two exponential terms describe the Doppler-induced phase evolution across OFDM symbols and the timing-induced phase slope across subcarriers. This diagonal abstraction omits residual inter-carrier interference, which remains present in the waveform-generated data. We use s to identify the two receiver-processing configurations used in this paper: s = 1 denotes the compensated NTN-1 condition and s = 2 the residual-impaired NTN-2 condition. For configuration s, let τ̂s and fˆd,RX,s denote the applied timing and residual-Doppler estimates. The remaining offsets after receiver processing are τpost,s = τ − τ̂s , fd,post,s = fd,pre − fˆd,RX,s . (3) Thus, transmitter pre-compensation removes the beamcommon satellite component, whereas receiver-side compensation estimates and removes the remaining location- and mobility-dependent Doppler. Let y denote the noisy received time-domain waveform and let yclean denote its matched counterpart, generated with the same waveform and channel realization but without additive noise. The receiver operator Rs {·} applies timing synchronization, optional residualDoppler compensation, and OFDM demodulation, producing the grid Y(s) = Rs {y}. Using synchronization parameters inferred from y, the matched noise-free effective target is clean Heff } ⊘ X, s = Rs {y
fd,common
fd,sat
UE
θUE
BC
θBC
Beam footprint
Fig. 1. Geometry of beam-common Doppler pre-compensation.
(1)
where fd,UE is the signed Doppler contribution caused by UE mobility [2]. Equivalently, transmitter pre-compensation maps the complex baseband waveform x(t) into √ xpc (t) = x(t)e−j2πfd,common t , where t denotes time and j = −1. To illustrate how the residual offsets appear on the OFDM grid, let i index subcarriers and k index OFDM symbols. The corresponding baseband subcarrier frequency and symbol reference time are fi and tk . A diagonal approximation of the received resource element before receiver compensation is pre can Yi,k = Hi,k Xi,k ej2πfd,pre tk e−j2πfi τ + Zi,k ,
LEO satellite
(4)
where X is the transmitted resource grid and ⊘ denotes element-wise division over its occupied resource elements. Unlike the auxiliary canonical (perfect) channel Hcan , e.g., used in [6] as target label, this target is expressed in the receiver reference frame after the selected synchronization and compensation operations. It is a one-tap equalization target; when residual ICI is non-negligible, it is conditioned on the transmitted grid and should not be interpreted as a dataindependent physical channel response. Let P denote the set of Demodulation Reference Signal (DM-RS) pilot positions. Because Xi,k is known at these (s) LS positions, the pilot-domain LS estimate is Ĥi,k = Yi,k /Xi,k for (i, k) ∈ P. A learned estimator receives this sparse grid, its linear interpolation, or model-specific pilot/data-aided features, and implements Ĥeff s = fΘ (Hin , u),
(5)
where fΘ denotes the estimator parameterized by Θ, Hin its channel input, and u auxiliary pilots, masks, or side information. The datasets in Section IV instantiate two choices of Rs . NTN-1 applies joint Physical Downlink Shared Channel (PDSCH) timing–frequency alignment, followed by independent time–frequency estimation and compensation of the residual Doppler fˆd,RX,1 . This produces a compensated input and matched effective target. NTN-2 retains the same transmitter pre-compensation and timing alignment but sets fˆd,RX,2 = 0. Residual phase evolution then remains in both its input and effective target. This pairing isolates receiver Doppler processing from channel and noise randomness. III. E VALUATION M ETHODOLOGY AND E STIMATORS Each model is trained independently from random initialization for NTN-1 and NTN-2. All noisy SNR versions of one underlying channel realization are assigned to the same partition. Otherwise, closely related versions of one channel could appear in both training and testing, causing data leakage and an optimistic generalization result. Test samples are therefore excluded from model training, normalization, early stopping, and hyperparameter selection. Inputs follow the original estimator design and complex grids use separate real and imaginary channels. Table I shows which estimators require interpolation or receiver-side information. This matters for deployment: sparse-
TABLE I I NPUTS OF THE COMPARED DL ESTIMATORS . Model SRCNN, ChannelNet, EDSR AttRNet ProEsNet LSiDNN-48 CE-ViT HELENA-MHSA, HELENA MDELAN-SISO
Channel input
Side information
LS+LI
None
Sparse LS Sparse LS LS+LI Sparse LS PA-LS+DA-LS+ positional encoding
None None SNR, Doppler, delay None Practical noise estimate
LS models avoid full-grid preprocessing, whereas CE-ViT and the model-driven method require additional inputs. The conventional references are LS with two-dimensional linear interpolation (LS+LI) [7], the practical 5G NR DM-RS estimator, and three LMMSE implementations [8]. Let h be the vectorized effective target, hp its pilot coefficients, and ĥLS p the noisy pilot observations. The implemented affine LMMSE estimator is −1 LS ĥp − hp , (6) ĥ = h + Rhp Rpp + (σ̂ 2 + λα)I where Np is the number of pilot coefficients. The fitting partition supplies the means and covariances and the SNRconditioned pilot-error variance σ̂ 2 , measured between noisy pilot LS observations and matched noise-free pilot labels. Moreover, λ = 10−4 and α = tr(Rpp )/Np . The linear system is solved without explicitly forming the inverse. Because LMMSE depends on second-order channel statistics, its performance also depends on how channels with different multipath structures are grouped. The datasets contain the standardized NTN-TDL-A–D profiles, where each profile defines a different tap-delay and power structure and the separately varied delay spread scales that structure [1]. We therefore evaluate three levels of statistical knowledge. Global LMMSE fits one covariance across all training and validation profiles and combines it with 11 SNR-conditioned pilot-error variances. Profile-aware LMMSE fits one covariance for each A–D profile, pooling its two delay-spread settings. Combining these four covariances with 11 SNR-conditioned pilot-error variances yields 44 filters, selected using the true profile and nominal SNR. Test-fitted profile-aware LMMSE additionally fits these statistics on the complete dataset, including test labels. It is non-deployable and is included only as an empirical linear reference, not as a theoretical bound. The practical NR estimator performs CDM-aware pilot processing, averaging, and full-grid interpolation [16], [17]. The DL comparison covers convolutional super-resolution (SRCNN and ChannelNet) [5], residual and attention refinement (EDSR, AttRNet, and ProEsNet) [9]–[11], and the fully connected LSiDNN [12]. Following the implementation details from [6], ChannelNet uses a frozen SRCNN front end and a 32-filter DnCNN; EDSR uses 32 filters and 16 residual blocks; AttRNet uses 32-filter attention-residual convolutions; and LSiDNN has 48 hidden neurons [6]. All convolutional models preserve the 612×14 grid without learned upsampling.
TABLE II PAIRED NTN DATASET CONFIGURATION . Parameter
NTN-1
NTN-2
Carrier/grid 2 GHz; 51 RB; 30 kHz SCS; normal CP Grid size 612 × 14 complex resource elements Link/antennas Downlink PDSCH; SISO Waveform symbols Full-grid QPSK DM-RS Type A; pos. 2; add. pos. 1; config. type 2 Channel profiles NTN-TDL-A, B, C, and D Delay spread A/B 30 or 100 ns Delay spread C/D 5 or 30 ns Satellite altitude 600 km UE elevation/speed 10–89◦ ; 0–120 km/h Pre-RX residual Doppler 150–900 Hz; beam offset up to 1.5◦ SNR/examples 0:2:20 dB; 11,264 per dataset PDSCH alignment Joint time–frequency TX pre-compensation Enabled Enabled RX Doppler compensation Independent time–frequency Disabled Primary target Compensated Residual-impaired
CE-ViT combines transformer blocks, channel metadata, and a transposed-convolution reconstruction layer [13]. We use this name because the original PD-CEViT pilot-design module is omitted; all estimators share the fixed NR DM-RS configuration. HELENA combines shallow convolution, compact multi-head self-attention, and squeeze-and-excitation, while HELENA-MHSA removes the latter block [6]. MDELANSISO follows the NTN channel-estimation design in [15], replacing the residual blocks of the communication-informed MDX receiver [14] with two MDELAN blocks. It retains the MDX preprocessing: interpolated PA-LS and a practical noise estimate yield per-RE SISO MMSE symbol estimates, from which DA-LS is computed. The blocks refine PA-LS, DA-LS, and positional inputs to produce the full-grid channel estimate. IV. NTN DATASET C REATION The paired datasets implement the two receiver operators R1 and R2 defined in Section II. They are generated at waveform level using MATLAB 5G Toolbox and Satellite Communications Toolbox, following the NR NTN PDSCH processing chain1 . Table II shows that propagation, numerology, pilots, and scenario sampling are common; receiver-side residual-Doppler compensation is the controlled difference. Each dataset contains 256 examples per SNR and NTNTDL profile, i.e., 4×11×256 = 11,264 samples. Channel realizations are balanced across the A–D multipath families, their applicable delay spreads, and three UE-speed, elevation, and residual-Doppler bins. Per profile, 179/38/39 realizations are assigned to training/validation/testing before generating their SNR variants, yielding 7,876/1,672/1,716 samples. Identical waveform, channel, and noise seeds make the two receiver conditions sample-wise comparable. SNR denotes the nominal AWGN injection level after normalized waveform/channel scaling; the dataset is intended for estimator comparison rather than satellite link-budget evaluation. Both conditions apply beam-common waveform precompensation to obtain xpc (t) and estimate τ̂s through 1 https://www.mathworks.com/help/satcom/ug/nr-ntn-pdsch-throughput. html
LS+LI 5G NR DM-RS estimator
Test-fitted PA-LMMSE HELENA
CE-ViT AttRNet
NTN-1: RX Doppler compensated
HELENA-MHSA MDELAN-SISO
NTN-2: residual impaired
0
NMSE (dB)
−5 −10 −15 −20 −25 −30 0
2
4
6
8
10
12
14
16
18
20
0
2
4
SNR (dB)
6
8
10
12
14
16
18
20
SNR (dB)
Fig. 2. Test NMSE versus SNR for selected estimators; Table III reports all results.
joint PDSCH frequency/timing alignment. NTN-1 estimates fˆd,RX,1 by combining cyclic-prefix fractional-offset and DMRS integer-subcarrier estimates [3], [4]. NTN-2 retains the alignment and timing tracking but sets fˆd,RX,2 = 0, as in Section II. All receiver-side timing and residual-Doppler estimates use noisy observations rather than true offsets. The practical baseline is nrChannelEstimate2 , configured with CDM lengths, averaging, interpolation, and noise estimation. Each MAT file stores the noisy grid, sparse/interpolated LS inputs, pilots and mask, practical DM-RS estimate, effective target, auxiliary canonical channel, split indices, SNR, scenario ID, and physical/receiver metadata. The evaluation assumes a known PDSCH/DM-RS configuration, isolating channel estimation from SSB detection, PBCH/control acquisition, and stateful initial synchronization. V. P ERFORMANCE E VALUATION
TABLE III SNR- AVERAGED TEST NMSE AND DL COMPLEXITY. Model
Params (×103 )
LS+LI 5G NR DM-RS G-LMMSE PA-LMMSE∗ TF-PA-LMMSE∗† (ref.) SRCNN ChannelNet EDSR ProEsNet AttRNet LSiDNN-48 CE-ViT HELENA-MHSA HELENA MDELAN-SISO
— — — — — 14.114 184.068 306.370 170.178 75.656 1662.240 880.370 114.162 116.290 1.396
FLOPs SNR-avg. NMSE (dB) (×109 ) NTN-1 NTN-2 — -8.000 -7.123 — -17.283 -11.832 — -21.921 -18.560 — -23.339 -19.331 — -23.658 -20.166 0.241 -18.668 -17.455 3.119 -21.020 -19.907 5.245 -21.288 -20.371 2.915 -21.227 -20.319 1.288 -21.334 -20.558 0.003 -17.946 -14.805 0.162 -21.068 -20.357 0.084 -20.739 -18.980 0.084 -21.856 -20.866 0.018 -18.308 -16.580
NTN-2 penalty (dB) ∆all +0.877 +5.451 +3.361 +4.008 +3.492 +1.213 +1.113 +0.918 +0.908 +0.776 +3.141 +0.710 +1.759 +0.990 +1.728
∆H +1.757 +6.567 +5.386 +6.461 +5.663 +2.259 +2.392 +2.084 +2.124 +1.947 +4.720 +1.773 +3.419 +2.612 +2.786
Lower mean NMSE is better. NMSE colors are relative to HELENA: green is better, red is worse, and blue denotes HELENA. Penalty = NTN-2 − NTN-1; positive values indicate degradation and are uncolored.
A. Experimental Setup and Reproducibility For each NTN condition, accuracy evaluation uses all 1,716 held-out examples, with 156 samples at each SNR. Models use Adam, MSE, batch size 64, at most 500 epochs, and early stopping with patience 50 and best-validation restoration. The initial learning rate is 10−3 with plateau reduction, except for MDELAN, which uses 0.03 and residual-scale initialization 10−4 after training-only tuning. Each model/condition uses one fixed-seed training run; sub-dB gaps are therefore observed rankings rather than statistically established margins. Compatible Keras models are exported to ONNX, verified numerically, and optimized with TensorRT using a common procedure. Datasets, compiled models, and evaluation scripts will be released online3 . TensorRT engines are compiled natively with a 4 GiB workspace and batch size one, for one channel grid per call. The RTX PRO 4500 and Jetson Orin NX represent high-performance and power-constrained embedded accelerator classes, respectively; the latter has been demonstrated in orbit as a commercial off-the-shelf (COTS)
AI accelerator [18]. We use both platforms only to compare the same estimator across compute classes; the modeled downlink CE remains at the UE. Both platforms use FP16: TensorRT 10.16 on RTX and TensorRT 10.3 with the Orin 10 W profile4 . Across ten models and both NTN conditions, the absolute FP16–FP32 NMSE difference averages 0.0011 dB and never exceeds 0.0038 dB without changing rankings. We therefore report only FP16 deployment results. For each model/platform, three runs use 500 warm-ups and 5,000 measurements. P99 is computed per run and then averaged across the three runs. Prepared-input-to-channeloutput latency includes pinned host-to-device and device-tohost transfers, inference, and synchronization, but excludes input construction and the remaining receiver chain. Power is sampled every 20 ms during workloads of at least 5 s. B. Channel Estimation Accuracy Fig. 2 first shows the per-SNR behavior. The NTN-2 curves flatten earlier than their NTN-1 counterparts, indicating that
2 https://www.mathworks.com/help/5g/ref/nrchannelestimate.html 3 https://github.com/miguelhdo/CE NTN Performance Evaluation
4 https://developer.nvidia.com/embedded/jetpack-sdk-62
TABLE IV NTN-1 FP16 DEPLOYMENT MEASUREMENTS AND HELENA- RELATIVE DELTAS . RTX PRO 4500 Blackwell, FP16 Jetson Orin NX, FP16, 10 W Mean±std. P99 Energy±std. ∆THEL ∆EHEL Mean±std. P99 Energy±std. ∆THEL ∆EHEL (ms) (ms) (mJ/sample) (%) (%) (ms) (ms) (mJ/sample) (%) (%) SRCNN 0.0361 ± 0.0006 0.0425 5.989 ± 0.059 −32.0 +5.9 0.6444 ± 0.0177 1.1717 4.110 ± 0.043 −1.5 −9.6 ChannelNet 0.1077 ± 0.0004 0.1120 19.641 ± 0.511 +103.0 +247.3 1.5319 ± 0.1242 2.2807 14.925 ± 1.467 +134.2 +228.3 EDSR 0.1528 ± 0.0008 0.1682 30.364 ± 0.322 +187.9 +436.9 2.1463 ± 0.1110 3.0533 22.373 ± 2.018 +228.1 +392.2 ProEsNet 0.2951 ± 0.0002 0.3016 54.492 ± 0.337 +456.1 +863.6 5.9873 ± 0.3372 6.9292 60.100 ± 6.155 +815.4 +1222.1 AttRNet 0.0754 ± 0.0011 0.0817 14.245 ± 0.664 +42.0 +151.9 1.1209 ± 0.0668 1.7964 9.772 ± 0.970 +71.4 +115.0 LSiDNN-48 0.0257 ± 0.0003 0.0290 2.961 ± 0.029 −51.5 −47.6 0.4075 ± 0.0048 2.4303 2.968 ± 0.088 −37.7 −34.7 CE-ViT 0.0967 ± 0.0009 0.1049 8.143 ± 0.375 +82.1 +44.0 0.8273 ± 0.0364 1.6245 5.236 ± 0.832 +26.5 +15.2 HELENA-MHSA 0.0478 ± 0.0004 0.0520 5.164 ± 0.126 −10.0 −8.7 0.6102 ± 0.0009 2.0781 4.057 ± 0.039 −6.7 −10.8 HELENA 0.0531 ± 0.0001 0.0595 5.655 ± 0.138 0.0 0.0 0.6541 ± 0.0230 1.7994 4.546 ± 0.210 0.0 0.0 MDELAN-SISO 0.1067 ± 0.0012 0.1136 8.999 ± 0.097 +101.0 +59.1 1.0348 ± 0.0795 2.1065 8.209 ± 1.589 +58.2 +80.6 Model
Green/red P99 cells meet/exceed the 0.5 ms budget. Positive ∆THEL and ∆EHEL denote higher mean latency and energy than HELENA. Energy/sample is baseline-inclusive workload energy divided by completed estimates. Mean±std. values are over three runs. RTX energy covers GPU-board power, whereas Orin energy covers complete-module input; energy values should therefore be compared within each platform.
NTN-1 SNR-averaged NMSE (dB)
Pareto frontier
RTX PRO 4500, FP16 −17
Jetson Orin NX, FP16, 10 W
LSiDNN
−18
LSiDNN SRCNN
SRCNN
−19 −20
0.5 ms budget
MDELAN H-MHSA
MDELAN
ChannelNet ProEsNet
AttRNet
−21
ChannelNet AttRNet
−22 HELENA
10
CE-ViT
EDSR
−1
P99 latency (ms, log scale)
H-MHSA
CE-ViT
10
HELENA EDSR
0
ProEsNet
101
P99 latency (ms, log scale)
Fig. 3. NTN-1 accuracy–tail-latency trade-off for FP16 inference.
residual impairments become dominant as AWGN decreases. HELENA remains among the most accurate DL estimators across both conditions, whereas the test-fitted LMMSE improves more strongly at high SNR in NTN-1 but exhibits an earlier error floor in NTN-2. These trends motivate the aggregate and paired-condition comparisons below. Table III reports the arithmetic mean of the eleven perSNR NMSE values expressed in dB over 0–20 dB. The paired differences, ∆all over 0–20 dB and ∆H over 12–20 dB, are computed as NTN-2 minus NTN-1. Positive values indicate degradation with residual Doppler. Percentages and factors express the corresponding linear-scale NMSE changes. HELENA obtains the lowest observed mean NMSE among DL estimators: −21.856 dB in NTN-1 and −20.866 dB in NTN-2. Relative to AttRNet, it reduces linear-scale NMSE by 11.3/6.8% (0.52/0.31 dB), while the reductions relative to CE-ViT are 16.6/11.1% (0.79/0.51 dB) in NTN-1/NTN-2. Channel recalibration reduces NMSE relative to HELENAMHSA by 22.7/35.2%, supporting its greater benefit with residual Doppler. HELENA also achieves 55.8/62.7% lower linear-scale NMSE than the NTN-specific MDELAN-SISO. The statistical LMMSE baselines exhibit a different, condition-dependent trend. In NTN-1, the global and profileaware variants outperform HELENA by 0.07 dB (1.5%) and
1.48 dB (28.9%), respectively. Test-fitted LMMSE is also 1.80 dB (34.0%) better than HELENA. This ordering reverses in NTN-2, where HELENA reduces linear-scale NMSE by 41.2%, 29.8%, and 14.9% relative to global, profile-aware, and test-fitted LMMSE, respectively. Thus, the test-fitted variant is an empirical linear reference, not a theoretical upper bound. Moving from NTN-1 to NTN-2 increases HELENA’s fullrange NMSE by 0.990 dB, corresponding to 1.26× higher linear-scale NMSE. AttRNet and CE-ViT exhibit smaller increases of 0.776 dB (1.20×) and 0.710 dB (1.18×), respectively, although HELENA retains the lowest absolute NMSE in both conditions. The practical NR estimator and global LMMSE are more sensitive, degrading by 5.451 dB (3.51×) and 3.361 dB (2.17×). At high SNR, HELENA’s penalty rises to 2.612 dB (1.83×), confirming the stronger influence of residual impairments as AWGN decreases. C. Deployment Performance Because the NTN-1 and NTN-2 models share architectures and tensor dimensions and differ only in their learned weights, Table IV reports deployment measurements only for NTN-1. At 30 kHz Subcarrier Spacing (SCS), one slot lasts 0.5 ms. Following the three-TTI receiver allowance adopted in HELENA [6], based on [17], we allocate one TTI to CE. Thus,
Tmax = 0.5 ms is an estimator-call evaluation budget, not a standardized complete-receiver deadline. On RTX, HELENA reaches 0.0531 ms mean latency and 0.0595 ms P99, 88.1% below the budget. It is 1.42×/1.82× faster than AttRNet/CE-ViT and uses 60.3/30.6% less energy. HELENA-MHSA saves 10.0% latency and 8.7% energy, but HELENA reduces its linear-scale NMSE by 22.7% (1.12 dB). On Orin, HELENA reaches 0.6541 ± 0.0230 ms mean latency and 1.7994 ms P99. It is 1.71×/1.26× faster than AttRNet/CE-ViT and uses 53.5/13.2% less energy. Its P99 nearly matches AttRNet’s but is 10.8% higher than CE-ViT’s. Fig. 3 shows HELENA on both platform frontiers, with the highest NTN-1 accuracy among the plotted models. VI. D ISCUSSION AND P RACTICAL I MPLICATIONS The results show that terrestrial channel-estimation architectures can remain effective after NTN retraining. In particular, the unchanged HELENA architecture achieves the best observed DL accuracy in both evaluated NTN conditions. This supports its robustness under supervised NTN adaptation, while zero-shot transfer from a TN-trained model remains to be evaluated. The degradation of every DL estimator in NTN-2 also shows that learning does not remove the benefit of receiver-side Doppler compensation. Although LMMSE is competitive with accurate channel statistics, acquiring and tracking them can be impractical; HELENA requires neither profile labels nor covariance matrices during inference. Deployment suitability depends strongly on the available computing resources. On RTX, HELENA provides the best DL accuracy while satisfying the adopted CE latency budget. On the 10 W Orin NX, no evaluated model meets this budget, including lightweight alternatives. Meeting this budget on regenerative or other power-constrained platforms therefore requires further advances, such as quantization, operator fusion, hardware-aware architectures, or FPGA acceleration. The MDELAN-SISO results also confirm that low parameter and FLOP counts do not guarantee low deployed latency or energy, making measurements on the target platform essential. These conclusions apply to the evaluated SISO, 600-km LEO, 2-GHz, 30-kHz SCS, and DM-RS configuration. Linklevel evaluation is still required to determine how the observed NMSE gains translate into BLER and throughput. VII. C ONCLUSION AND F UTURE W ORK We evaluated HELENA and nine DL estimators under receiver-compensated and residual-impaired 5G NR NTN conditions. After NTN retraining, the unchanged HELENA architecture achieves the lowest observed mean DL NMSE and a favorable latency–energy trade-off. It satisfies the adopted CE latency budget on RTX, whereas no evaluated model meets the P99 budget on the 10 W Orin NX. The results also show that residual Doppler remains an important source of high-SNR degradation and that analytical model complexity does not reliably predict deployed performance. Future work will evaluate link-level MIMO performance, terrestrial-toNTN transfer learning, and hardware-aware acceleration for constrained receivers.
ACKNOWLEDGMENTS This work was supported by imec.icon RAPIDNESS, cofinanced by imec and Flanders Innovation and Entrepreneurship (HBC.2024.0772). R EFERENCES [1] 3GPP, “Study on new radio (nr) to support non-terrestrial networks,” 3rd Generation Partnership Project, Tech. Rep. TR 38.811, 2020, version 15.4.0. [2] A. K. Meshram, S. Kumar, J. Querol et al., “Reduced complexity initial synchronization for 5G NR multibeam LEO-based non-terrestrial networks,” IEEE Open Journal of the Communications Society, vol. 6, pp. 1528–1551, 2025. [3] J.-J. van de Beek, M. Sandell, and P. O. Börjesson, “ML estimation of time and frequency offset in OFDM systems,” IEEE Transactions on Signal Processing, vol. 45, no. 7, pp. 1800–1805, 1997. [4] C.-C. Chen, Y.-S. Huang, C.-L. Tu et al., “Hardware-efficient doppler estimation and compensation in PDSCH for 5G non-terrestrial networks,” in 2025 IEEE 16th International Conference on ASIC (ASICON), 2025, pp. 1–4. [5] M. Soltani, V. Pourahmadi, A. Mirzaei et al., “Deep learning-based channel estimation,” IEEE Communications Letters, vol. 23, no. 4, pp. 652–655, 2019. [6] M. Camelo Botero, E. A. Beyazıt, N. Slamnik-Kriještorac et al., “Helena: High-efficiency learning-based channel estimation using dual neural attention,” in 2026 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit), 2026, pp. 40–45, málaga, Spain. [7] S. J. Lee, “On the training of mimo-ofdm channels with least square channel estimation and linear interpolation,” IEEE Communications Letters, vol. 12, no. 2, pp. 100–102, 2008. [8] O. Edfors, M. Sandell, J.-J. van de Beek et al., “Ofdm channel estimation by singular value decomposition,” IEEE Transactions on Communications, vol. 46, no. 7, pp. 931–939, 1998. [9] D. Maruyama, K. Kanai, and J. Katto, “Performance evaluations of channel estimation using deep-learning based super-resolution,” in 2021 IEEE 18th Annual Consumer Communications and Networking Conference (CCNC), 2021, pp. 1–6. [10] W. Gao, W. Zhang, L. Liu et al., “Deep residual learning with attention mechanism for ofdm channel estimation,” IEEE Wireless Communications Letters, vol. 14, no. 2, pp. 250–254, 2025. [11] Y. Zhang, J. Hou, and H. Liu, “Deep learning based fully progressive image super-resolution scheme for channel estimation in ofdm systems,” IEEE Transactions on Vehicular Technology, vol. 73, no. 6, pp. 9021– 9025, 2024. [12] A. Sharma, S. A. U. Haq, and S. J. Darak, “Low complexity deep learning augmented wireless channel estimation for pilot-based ofdm on zynq system on chip,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 71, no. 5, pp. 2334–2347, 2024. [13] F. Liu, P. Jiang, J. Zhang et al., “Pd-cevit: A novel pilot pattern design and channel estimation network for ofdm systems,” IEEE Transactions on Communications, vol. 73, no. 6, pp. 4363–4377, June 2025. [14] M. Abdollahpour, M. Bertuletti, Y. Zhang et al., “A compute&memory efficient model-driven neural 5g receiver for edge ai-assisted ran,” in IEEE Global Communications Conference (GLOBECOM), 2025, pp. 5248–5253. [15] M. Abdollahpour, B. De Filippo, C. Amatetti et al., “On-board ai-based channel estimation for leo ntns,” arXiv preprint arXiv:2607.15127, 2026, accepted at IEEE SPAWC 2026. [16] Y. Li, “Pilot-symbol-aided channel estimation for ofdm in wireless systems,” IEEE Transactions on Vehicular Technology, vol. 49, no. 4, pp. 1207–1215, 2000. [17] S. A. Damjancevic, E. Matus, D. Utyansky et al., “Channel estimation for advanced 5g/6g use cases on a vector digital signal processor,” IEEE Open Journal of Circuits and Systems, vol. 2, pp. 265–277, 2021. [18] D. Won, K. Hong, D. Lee et al., “In-orbit validation of a heterogeneous COTS architecture for onboard AI: System-level operational vulnerabilities and failure pathways,” International Journal of Aeronautical and Space Sciences, vol. 27, no. 4, pp. 3642–3659, Jul. 2026.