PaperlayerView exact revision on arXiv

Source-linked AI summary

Digital Twin Enhanced Channel Twin for AI-Native CSI Inference: Generalizability and Scalability

Majumder Haider, Imtiaz Ahmed, Zoheb Hassan, Danda B. Rawat, Huaiyu Dai

arXiv:2609.27017v1eess.SP

TL;DR

Per-symbol calibrated ray tracing is too slow for microsecond latency budgets, and cross-environment channel-twin validation and unified-receiver evaluation remain gaps. The paper combines sparse ray-tracing anchors with Transformer interpolation, spatial transfer, and a unified neural receiver; it reports accurate sparse-anchor reconstruction, transfer to a nearby unseen environment, and the lowest BER among considered receiver schemes.

Abstract

from arXiv · show

Accurate channel state information (CSI) is critical for advanced multi-antenna wireless networks. While high-fidelity and site-specific ray-tracing (RT) equipped wireless digital twins can overcome overhead for CSI acquisition. However, computing deterministic, calibrated RT based CSI from a wireless digital twin for every orthogonal frequency-division multiplexing (OFDM) symbol violates the strict microsecond latency budgets of the 5G NR numerology. To overcome this computational bottleneck, this paper investigates three approaches within a calibrated 3D digital twin framework, namely (i) the generalization of the channel twin, (ii) the performance enhancement using a neural receiver, and (iii) a data-driven interpolation framework for scalability. We compute high-precision CSI for a sparse subset of temporal anchors and employ an attention-based Transformer to predict the remaining symbols. Unlike polynomial splines or sequential long short-term memory (LSTM) networks, the Transformer exploits a global receptive field to capture the non-linear multipath dynamics while enabling parallelizable, real-time inference. To achieve spatial scalability, we introduce channel twin generalization. By fine-tuning the 3D RT models alongside the Transformer, the framework leverages the CSI dataset of one location to infer the channel behavior of an unseen environment. Transfer learning further adapts the model to a new environment using only a small fraction of locally collected data. Simulation results demonstrate that the proposed architecture substantially outperforms the baseline interpolators, achieves robust spatial transferability, and lowers the bit error rate through the unified neural receiver. These results establish a scalable, environment-agnostic foundation of distributed channel twins for high-quality, low-overhead CSI acquisition in next-generation cellular networks.

I. INTRODUCTION

Accurate CSI supports massive-MIMO beamforming and spectral efficiency, but acquiring it in dynamic environments creates overhead that does not scale to anticipated 6G arrays and bandwidths. Prior work combines AI-based CSI methods with digital twins, while cross-environment validation and integration with unified neural receivers remain gaps.

A. Introductory Background

  • High-fidelity, real-time CSI supports directive beams and spectral efficiency in massive MIMO, with particular importance at blockage-sensitive sub-THz frequencies.CSI is also critical for integrated sensing and communication.Source: A. Introductory Background · p. 1
  • Legacy CSI estimation and feedback signaling overhead does not scale to anticipated 6G array sizes and bandwidths, motivating low-overhead AI-native approaches.Site-specific wireless digital twins can generate physics-grounded channel data for training these models.Source: A. Introductory Background · p. 1

B. Related Work

  • Prior research uses deep learning and ray-tracing digital twins to generate synthetic CSI, with reconstruction gains depending on faithful geometry and hardware modeling.Twin-generated CSI can support offline training, what-if analysis, network planning, and proactive optimization.Source: B. Related Work · pp. 1–2
  • Unified neural receivers replace separate signal-processing stages with joint learned processing, while tensor decomposition has been used to compress such architectures for latency and storage budgets.Related examples combine demodulation with channel or source decoding.Source: B. Related Work · p. 2
  • Existing digital-twin channel models are rarely validated across distinct propagation environments or integrated with unified neural receivers under real-time latency constraints.The identified gap also includes reconstructing full CSI from sparse high-precision anchors within 6G numerology budgets.Source: B. Related Work · p. 2

C. Motivation & Contribution

Per-symbol calibrated ray tracing exceeds the latency budget, while conventional interpolators may miss nonlinear multipath dynamics or incur sequential processing bottlenecks. The paper combines sparse ray-tracing anchors, Transformer interpolation, spatial transfer, and a unified neural receiver within a MISO-OFDM system model.

C. Motivation & Contribution

  • Per-symbol calibrated ray tracing violates microsecond latency budgets, and splines or sequential LSTMs struggle with nonlinear multipath dynamics or wide temporal gaps.The proposed alternative uses an attention-based Transformer with a global temporal receptive field for parallelizable interpolation from sparse anchors.Source: C. Motivation & Contribution · p. 2
  • Channel-twin generalization uses fine-tuned 3D ray-tracing models and AI to infer channel characteristics in unseen environments from field-collected CSI at a mapped location.The scene-to-scene transfer is intended to reduce the need for exhaustive site-specific data collection.Source: C. Motivation & Contribution · p. 2
  • A unified neural receiver jointly performs demodulation and decoding in one AI-driven block rather than separate processing stages.The contribution passage says this reduces processing overhead and over-the-air pilot transmission while improving robustness to nonlinear impairments.Source: C. Motivation & Contribution · p. 2

II. SYSTEM MODEL

  • The system model considers a time-varying, frequency-selective MISO-OFDM downlink, with multiple base-station antennas and a single-antenna user equipment.It uses linear precoding and a channel-dependent precoding factor.Source: §II · pp. 2–3
  • The channel twin supplies the base-station precoder's CSI, while the receiver uses pilot-based channel estimation and linear equalization before detection.The conventional receiver then demodulates equalized symbols and decodes the estimated coded bits in separate operations.Source: §II · p. 3

III. PROPOSED FRAMEWORK

The framework has three components for spatial generalization, receiver performance, and temporal scalability. Its system figure presents two scenes, their 3D ray-tracing models, and system parameters.

III. PROPOSED FRAMEWORK

  • The channel-twin framework combines spatial generalization, a joint demodulation-and-decoding neural receiver, and Transformer-based temporal interpolation.These components address spatial scalability, performance enhancement, and temporal scalability, respectively.Source: §III · p. 3
  • The framework's component list distinguishes spatial generalization from temporal interpolation and neural-receiver processing.The passage identifies these as separate parts of the proposed channel twin.Source: §III · p. 3
  • Figure 1 pairs satellite views of Scene 1 and Scene 2 with their corresponding 3D ray-tracing models and lists the system parameters.Panels (a) and (c) show the views, panels (b) and (d) the models, and panel (e) the parameters.Source: Figure 1 · p. 3

A. Generalization of Channel Twin

The channel twin combines a calibrated 3D ray-tracing model with an AI model that maps its output toward the real-world channel. Its generalization process uses shared propagation structure between nearby environments, with calibration performed offline.

A. Generalization of Channel Twin

  • The channel twin is the calibrated 3D RT model together with an AI model that maps its output close to the corresponding real-world channel.The paper tunes radio-material properties and key RT parameters offline; continuous closed-loop refinement is beyond its present scope.Source: A. Generalization of Channel Twin · pp. 3–4
  • Generalization is scoped to nearby environments with shared morphology, where spatially correlated electromagnetic fields yield coupled propagation paths and delay profiles.The two measured Northern Virginia scenes satisfy this condition.Source: A. Generalization of Channel Twin · p. 4
  • The three phases are data acquisition and modeling, ray-tracing and AI-model training, and inference and validation against the target-scene ground truth.Training uses real CSI and synthetic datasets from fine-tuned RT models; inference uses source-scene data to predict target-scene CSI.Source: A. Generalization of Channel Twin · p. 4
  • The framework predicts CSI at a target location from a known location using RT and AI to capture correlations arising from environmental physics and wireless propagation.The paper evaluates target-scene CSI precision across three test cases.Source: A. Generalization of Channel Twin · p. 4

B. Performance Enhancement Using Neural Receiver

The proposed neural receiver replaces separate demodulation and decoding blocks with a data-driven block that jointly maps equalized symbols to information bits. The design uses a lightweight feed-forward network suited to UE computational constraints.

B. Performance Enhancement Using Neural Receiver

  • The conventional receiver's separate channel-estimation and equalization stages can compound errors under multipath fading, inter-antenna interference, and hardware non-linearities, degrading decoder soft bits and increasing BER.The conventional pipeline assumes a linear, Gaussian channel and perfect synchronization.Source: B. Performance Enhancement Using… · p. 5
  • The neural receiver jointly performs demodulation and channel decoding, mapping the equalized symbol grid directly to estimated information bits.It removes the sequential hand-off between a separate demapper and Hamming decoder.Source: B. Performance Enhancement Using… · p. 5
  • Trained on diverse channel realizations, the neural receiver learns a soft mapping from equalized symbols to information bits, jointly exploiting code structure and constellation geometry while compensating for residual impairments.The paper presents this joint learning as a way to avoid a hard, sequential demapper-to-decoder hand-off.Source: B. Performance Enhancement Using… · p. 5
  • The receiver is a lightweight fully connected feed-forward network, chosen for the UE's limited computational resources.Its input is the equalized symbol grid and its output is the estimated information bits.Source: B. Performance Enhancement Using… · p. 5

C. Scalability of Channel Twin

To reduce the cost of computing CSI for every OFDM symbol, the framework computes high-precision CSI at sparse temporal anchors and interpolates the remaining symbols. A Transformer processes the anchor context and reconstructs the subframe in a parallel pass.

C. Scalability of Channel Twin

  • For 30 kHz numerology, each subframe has 28 OFDM symbols in two 14-symbol slots; CSI is computed by RT and AI for selected estimation symbols and interpolated for the rest.These selected symbols are also called anchor symbols or temporal anchors.Source: C. Scalability of Channel Twin · p. 5, Figure 4 · p. 6
  • The RT impulse response represents multipath components through their complex gains and delays, and the resulting anchor CSI supplies the input for AI-enhanced interpolation.The paper defines the path amplitude, delay, operating frequency, path count, and Dirac delta function in the displayed expression.Source: C. Scalability of Channel Twin · p. 5
  • At 751 MHz and speeds up to 100 km/h, the channel coherence time is approximately 6 ms, exceeding the 1 ms subframe duration and supporting interpolation across its 28 symbols.The maximum Doppler shift is approximately 70 Hz.Source: C. Scalability of Channel Twin · p. 5
  • The Transformer encoder contextualizes the subframe by attending across tokens in parallel, while its decoder uses the anchor context to reconstruct all symbol positions jointly.Because all anchors are available in advance, the implementation relaxes the decoder's future-position masking and reconstructs 28 high-resolution CSI matrices in one parallel pass.Source: C. Scalability of Channel Twin · p. 6

D. Complexity Analysis and Real-Time Feasibility

The scheme confines deterministic RT computation to estimation symbols, uses a tunable ray budget with U-Net refinement, and predicts the remaining CSI with a parallel Transformer. Its per-subframe inference comprises seven RT executions, seven U-Net inferences, and one Transformer inference within 1 ms.

D. Complexity Analysis and Real-Time Feasibility

  • RT cost scales as O(N_ray β), while the fixed-layer U-Net costs O(L_u κ C^2 η), linear in CIR length η; together, estimation cost scales with N_est(N_ray β + L_u κ C^2 η).Reducing the ray count lowers per-symbol RT cost, while the U-Net refines the resulting low-fidelity CIR.Source: D. Complexity Analysis and Real-Time… · pp. 6–7
  • With N_sym = 28 fixed by numerology, the Transformer processes the entire subframe in a single parallel pass with sequential depth O(1), and its quadratic term is modest.Source: D. Complexity Analysis and Real-Time… · p. 7
  • Network training costs are incurred offline and therefore do not affect the per-subframe latency budget.The training costs accumulate over samples and epochs as multiples of each network's forward-pass cost.Source: D. Complexity Analysis and Real-Time… · p. 7
  • Seven RT executions, seven U-Net inferences, and one Transformer inference fit within the 1 ms subframe duration.The scheme uses seven estimation symbols out of 28, reducing the RT invocation rate by a factor of four.Source: D. Complexity Analysis and Real-Time… · p. 7

IV. SIMULATION RESULTS

The evaluation uses BER as its principal metric and relates received SNR and spectral efficiency to beamforming quality. Measurements use a 751 MHz cellular deployment, with a lightweight coded link and specified channel-data settings.

A. Parameter Specifications

  • BER is the principal performance metric, while received SNR depends on alignment between the precoder and the actual channel estimated from CSI.The stated SE relation is SE = log2(1 + γ_R).Source: A. Parameter Specifications · p. 7
  • The reported spectral efficiency is Shannon SE at achieved received SNR, and its cross-scheme gaps quantify precoder alignment under a common channel scaling.More accurate CSI supports better precoder alignment, higher effective beamforming gain, and higher attainable rate.Source: A. Parameter Specifications · pp. 7–8
  • The measurement setup uses a 751 MHz live deployment, 25 mi/h UE speed, and 105 CIR snapshots per scene, split 80% for training and 20% for testing.The two scenes are evaluated with 4 transmit antennas, one receive antenna, 128 subcarriers, and QPSK with a ZF equalizer.Source: A. Parameter Specifications · pp. 7–8

B. Generalization of Channel Twin

The HF-DT framework transfers a learned RT-to-real CIR mapping from one drive-test scene to another. In the unseen scene, BER and spectral-efficiency curves track the matched-scene configuration, although both show a performance gap.

B. Generalization of Channel Twin

  • The Actual-CIR-Scene1 curve assumes perfect real-world CSI unavailable to the base station, so it is a theoretical BER lower bound rather than a realizable benchmark.It attains the best BER across the SNR range.Source: B. Generalization of Channel Twin · p. 8
  • The matched configuration trains and deploys the U-Net mapping in Scene 1, using RT-generated Scene 1 CIR to estimate the channel for precoding.The test uses the real-world Scene 1 channel for BER evaluation.Source: B. Generalization of Channel Twin · p. 9
  • The generalized configuration trains the U-Net RT-to-real CIR mapping on Scene 2 only, then uses Scene 1 RT CIR to infer a channel for Scene 1 precoding.The test compares Scene 1 evaluation with matched Scene 1 training and generalization from Scene 2.Source: B. Generalization of Channel Twin · pp. 8–9
  • In the unseen Scene 1, the Scene 2-trained configuration tracks the matched model in BER and spectral efficiency, but with a performance gap.The genie-aided perfect-channel benchmark has the best BER across SNR, while matched HF-DT achieves near-optimal beamforming.Source: B. Generalization of Channel Twin · pp. 8–9, Figure 5 · p. 8, Figure 6 · p. 8

C. Transfer Learning Performance

Fine-tuning on a fraction of target-scene data improves BER and spectral efficiency relative to using no target-scene fine-tuning. Using 30% rather than 10% of Scene 1 data brings performance closer to the jointly trained reference.

C. Transfer Learning Performance

  • Fine-tuning improves BER toward the jointly trained reference, and using 30% of Scene 1 data yields a closer match than using 10%.The jointly trained reference uses combined Scene 1 and Scene 2 datasets and is tested on Scene 1.Source: C. Transfer Learning Performance · p. 9
  • The transfer-learning procedure pretrains on Scene 2 and fine-tunes on either 10% or 30% of Scene 1 data before Scene 1 evaluation.The authors report that this uses a small fraction of locally collected data and reduces empirical data-collection burden.Source: C. Transfer Learning Performance · p. 9
  • In spectral efficiency, both fine-tuned configurations approach the jointly trained reference as SNR increases, with 30% data outperforming 10%.This ordering mirrors the BER results.Source: C. Transfer Learning Performance · p. 9

D. Neural Receiver Performance

The Transformer-based interpolator achieves substantially lower NMSE than the spline and LSTM baselines, while the unified neural receiver achieves the lowest BER across the tested SNR range. The section also describes the interpolation task and the attention mechanism used to predict unobserved symbols.

D. Neural Receiver Performance

  • RT-NN-NR achieves the lowest BER across the SNR range, and both neural-receiver configurations outperform the RT-NN-Task2 baseline.The RT-NN-NR versus Synthetic-NR gap highlights the value of HF-DT channel knowledge for precoder design compared with a Rayleigh assumption.Source: D. Neural Receiver Performance · p. 10

E. Scalability and Interpolation Performance

  • The interpolation task uses CIRs from estimation symbols to predict CIRs for the remaining symbols within a subframe.The prediction follows the time-slot structure described in Fig. 4.Source: E. Scalability and Interpolation… · p. 10
  • Transformer interpolation attains −17.2584 dB NMSE, approximately 29 dB better than the best baseline.The Transformer is the only scheme below 0 dB; the HF-DT-anchor LSTM and spline baselines attain 11.7770 dB and 12.5592 dB, respectively.Source: E. Scalability and Interpolation… · p. 10, Table I · p. 10
  • The Transformer and LSTM use comparable trainable-parameter counts and identical anchor data, loss, and training protocol, so their NMSE difference reflects architecture rather than training budget or model capacity.The comparison is between global attention and sequential recurrence.Source: E. Scalability and Interpolation… · p. 10
  • Self-attention relates every sparse anchor to every missing symbol across the time horizon in parallel, adapting anchor relevance to instantaneous channel conditions.The passage contrasts this global processing with spline smoothness assumptions and LSTM sequential processing, and states that Transformer prediction meets microsecond latency requirements for 30 kHz numerology.Source: E. Scalability and Interpolation… · p. 10

V. CONCLUSIONS

The conclusion presents an AI-empowered channel-twin framework for accurate CSI reconstruction from sparse temporal anchors within the 5G NR subframe budget. It reports spatial transfer, neural-receiver, and interpolation results, while identifying broader environment validation as future work.

V. CONCLUSIONS

  • The HF-DT framework uses a U-Net to learn the relationship between RT-generated and real-world CIRs, reconstructing channels from sparse temporal anchors within the 1 ms subframe budget.The complexity analysis establishes feasibility for 5G NR 30 kHz numerology.Source: §V · p. 10
  • A model trained in one mapped environment transferred successfully to a nearby unseen environment with similar morphology, and fine-tuning on a small fraction of target data approached the jointly trained reference.The conclusion identifies validation across more environment pairs, the reverse transfer direction, and morphologically distinct sites as future work.Source: §V · p. 11
  • The unified neural receiver achieved the lowest BER among considered schemes, reduced baseband processing overhead, and improved resilience to residual channel impairments.The improved channel estimates also corresponded to higher achievable spectral efficiency across the operating range.Source: §V · p. 11
  • The Transformer interpolator achieved markedly lower NMSE than the spline and LSTM baselines, supporting its scalability for real-time CSI prediction.Source: §V · p. 11
Loading 2609.27017v1…