Poster: Practical Cross-Band Channel Prediction for AI-RAN via Physics-Guided Deep Unfolding Ruiqi Kong, He Chen∗ , Xiaojun Lin
Department of Information Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China Email: {rqkong, he.chen, xjlin}@ie.cuhk.edu.hk Table 1: Comparison of channel prediction methods.
arXiv:2605.31279v1 [eess.SP] 29 May 2026
Abstract To make cross-band channel prediction practical for AI-native RAN, algorithms must generalize across diverse environments and support real-time inference. Existing approaches achieve one but not both. To bridge this gap, we introduce GUIDE, a physics-guided deep unfolding framework that embeds wireless channel physics into differentiable layers. Without retraining in unseen environments, GUIDE achieves 2.75× beamforming gain than the deep learning-based baseline FIRE with only a slight increase in inference time, and 1.39× beamforming gain than the strongest model-based baseline R2F2 while running over 1610× faster.
CCS Concepts • Networks → Wired access networks; Mobile networks; • Computing methodologies → Artificial intelligence.
Keywords AI-RAN, cross-band channel prediction, deep unfolding. ACM Reference Format: Ruiqi Kong, He Chen∗ , Xiaojun Lin . 2026. Poster: Practical Cross-Band Channel Prediction for AI-RAN via Physics-Guided Deep Unfolding. In The 24th Annual International Conference on Mobile Systems, Applications and Services (MobiSys Companion ’26), June 21–25, 2026, Cambridge, United Kingdom. ACM, New York, NY, USA, 2 pages. https://doi.org/10.1145/3812835.3814874
1
Introduction
Among the AI-native capabilities envisioned for AI-RAN, crossband channel prediction is particularly valuable for frequency division duplexing (FDD) cellular systems [2, 3], where downlink (DL) channel state information (CSI) acquisition typically relies on explicit user feedback [1]. In massive MIMO, this feedback incurs substantial bandwidth overhead and latency [2]. The key physical insight enabling feedback-free FDD is that the underlying multipath parameters are frequency-independent. If these parameters can be accurately estimated from the uplink (UL) channel, the DL channel at any frequency can be reconstructed directly. Existing approaches adopt two broad paradigms, each with distinct trade-offs, as summarised in Table 1. Model-based methods such as R2F2 [4] and Horcrux [2] explicitly exploit the physical *Corresponding author.
This work is licensed under a Creative Commons Attribution-NonCommercialNoDerivatives 4.0 International License. MobiSys Companion ’26, Cambridge, United Kingdom © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2711-5/2026/06 https://doi.org/10.1145/3812835.3814874
Property Env. generalization Fast inference Multipath-rich envs
R2F2
FIRE
HORCRUX
GUIDE
✓ × ×
× ✓ ×
✓ × ×
✓ ✓ ✓
structure: R2F2 iteratively estimates multipath parameters, offering strong interpretability and cross-environment transferability, yet its greedy search is computationally prohibitive and collapses under dense multipath; Horcrux improves robustness through sub-band neural estimators but retains a disjoint two-stage pipeline that is not jointly optimised, limiting accuracy and increasing latency. Black-box deep learning methods such as Fire [3] sidestep explicit modelling entirely, achieving fast inference by learning a direct UL-to-DL mapping, but at the cost of poor generalization; the learned mapping must be retrained whenever environments or system setups change. This motivates a model-driven deep learning approach that combines the physical interpretability and transferability of modelbased methods with the speed and end-to-end trainability of datadriven ones. GUIDE achieves this via deep unfolding: the iterative estimation of multipath parameters is mapped to multiple differentiable network layers with physics constraints embedded structurally. This makes GUIDE accurate, fast, and transferable to unseen environments without retraining the neural network itself.
2
GUIDE Design
GUIDE unfolds the iterative estimation of multipath parameters {distance 𝑑𝑛 , complex attenuation 𝑎𝑛 } into a stack of 𝐿 differentiable physics-guided layers trained end-to-end via backpropagation. Realizing this idea, however, raises two non-trivial challenges: (i) determining which physical constraints to be embedded rigidly versus left to the data, and (ii) ensuring stable gradient flow through layers after physical constraints embedding. The design of GUIDE is shaped by both, as shown in Fig. 1. The observed UL CSI matrix HUL is first processed by a multiscale CNN, capturing both local and global structure across subcarriers to produce a shared feature representation for all subsequent steps. Reliable initialization of multipath parameters is essential before unfolding: a poor starting point risks slow convergence or entrapment in local optima. A learned initializer layer therefore uses cross-attention to aggregate per-antenna features into 𝑁 path-specific estimates d0 , placing the solution close to the true parameter manifold. The unfolding stage then refines these estimates over 𝐿 iterations, directly confronting challenge (i): embedding too many constraints reduces model capacity under environment mismatch,
oss-Band Channel Prediction MobiSys 2025 MobiSys Companion ’26, June 21–25, 2026, Cambridge, United Kingdom
nfold
UL Channel HUL
ve estimation of multitack of L differentiable o-end via backpropagapeline.
Channel Feature Extractor Multi-Scale Initial Estimator
r
CIK (where I is the number of antennas) is ×K . We stack real and r (B, 2, I, K) and pass
P) ∈ RB×C×I×K
(2)
onvolutional branches to capture features at hannel-wise concatena-
d0 ∈ RN
Construct Steering Matrix Sl
×L layers dl+1 al+1
Physics-Guided Update Module
(a) Beamforming Gain.
(b) Speed–Accuracy Trade-off.
Figure 2: Performance comparison against baselines. LMMSE Estimator (closed-form)
â ∈ CN
DL Channel Reconstruction ■ Physics (fixed)
■ Learned (trainable)
2. DeepUnfold architecture. Blue blocks are Figure Figure 1: GUIDE architecture. Blue blocks are physics-derived tor physics-derived (differentiable but fixed-form); orange blocks (differentiable but fixed-form); orange are are learned neural modules. The dashed purple blocks box denotes the learned path-specific parameter unfolding loop. neuralL-layer modules. The dashed purple box denotes the 𝐿-layer ntion:
atten(F))
Ruiqi Kong, He Chen∗ , Xiaojun Lin
unfolding loop. (3)
e query vectors. Initial 3. Experimental Evaluation while embedding too few sacrifices cross-environment transferabilsed: 3.1. resolves Setup this tension by hard-coding only the steering ity. GUIDE RB×2×N (4) Simulation. We generate channels using Eq. (1) with
matrix construction, encoding propagation distance via complex N ∈ {2, 4, 8} paths, dn ∼ U(1, 200) m, an ∼ U(0, 1), ϕn ∼ exponentials, while delegating residual correctionand to aSNR lightweight U(−π, π), 20 MHz bandwidth, 64 subcarriers, , . . . , L, the system: learned∈module. Specifically, at 200K each simulated iteration samples; 𝑙, the steering ma[10, 25] dB. Training uses no matrix Sk ∈ CIK×N : S𝑙 is real-world data isanalytically used for training. trix constructed from current estimates d𝑙 , and rant ·û(ψk,n ) Hardware. We−implement DeepUnfold on WARP v3 the residual r𝑙 = HUL S𝑙 a𝑙 provides a physically interpretable π λi (5) SDRs with a 2/4-antenna base station. Uplink is measured error signal. The learned corrector then outputs refined estimates on channel 1 (2.412 GHz); downlink is predicted on channel nd û(ψ) is the unit di-, absorbing d𝑙+1 model mismatch without the under6 (2.437 GHz, ∆f = 25 MHz) and channelre-learning 11 (2.462 GHz, ∆f =physics. 50 MHz). Challenge (ii) is addressed by the analytical lying wave = p − Sk ak , where ak Metric. prediction accuracy = of S𝑙 is differentiability the steering since(dB): eachAcc entry ofChannel matrix: −10 log10of∥H − Ĥ∥2 /∥ Ĥ∥2 . exponentials, gradients propaa composition smooth complex learned correction: Baselines. FIRElayers [2], HORCRUX [3]. All gate cleanly throughR2F2 all 𝐿 [1], physics during backpropagation dUpdate(ψk , dk , rk , Tk ) baselines are reimplemented from scratch. without approximation. Once geometric parameters converge, the (6) 3.2. Results complex path attenuations â are recovered in closed form via the ation LMMSE estimator, which remains fully differentiable throughout 3.3. Discussion udes are estimated via training. Table 1 summarizes the key properties of all meth-
high beamforming gains across test cases. Although R2F2 can occasionally produce competitive estimates, its large variation suggests limited robustness under the evaluated conditions. FIRE and HORCRUX, on the other hand, fail to provide reliable beamforming improvement. Fig. 2(b) further reveals the speed-accuracy tradeoff across methods. GUIDE achieves the best of both regimes: it approaches the ground-truth beamforming gain while completing inference in under 4 ms, over 1610× faster than R2F2. Therefore, the proposed method provides a favorable balance between accuracy and efficiency, enabling a real-time AI-RAN capability for accurate channel inference and low-latency RAN optimization.
Layers
1
SH Lp
3
(7)
ods. DeepUnfold is the only system that simultaneously achieves physics grounding, cross-environment generalizaExperimental Evaluation tion, cross-frequency scalability, and end-to-end trainability.
ble layer, allowing graWe synthesize 80K training channels under the CDL-B non-line-ofand dL .
4. Conclusion sight (NLoS) model with 10 MHz bandwidth, 26 subcarriers, and We20] presented a physics-guided deep unSNR ∈ [10, dB. WeDeepUnfold, evaluate generalisation on over-the-air meafolding network for cross-band channel prediction in FDD L is reconstructed by surements collected in a complex office environment under NLoS systems. By unrolling iterative multipath parameter esticonditions using Xilinx ZC706layers SDRwith withembedded four antennas. mation into ajointly trainable physics Uplink d̂n +j ϕ̂n constraints, DeepUnfold state-of-the-art accuchannels are measured at 2.412achieves GHz and downlink channels are DL,i (8) racy across diverse environments and frequency bands withpredicted at 2.422 GHz. No retraining is performed on real-world data. All training and inference are conducted on an NVIDIA RTX 4070 4GPU. Three baselines are evaluated: R2F2 [4], FIRE [3], and HORCRUX [2], with beamforming gain as the metric. Fig. 2 compares the downstream beamforming performance and inference efficiency of the proposed method against representative baselines. Fig. 2(a) reports the beamforming gain across all methods in real-world measurements. GUIDE achieves a mean value of 5.20 dB, outperforming R2F2 (3.76 dB), FIRE (0.81 dB), and HORCRUX (0.34 dB) by +1.44 dB, +4.86 dB, and +4.39 dB, respectively, and approaching within 0.76 dB of the ground-truth (GT). The error bars in Fig. 2(a) further show that our method provides consistently
on
4
Conclusion
GUIDE achieves accurate and environment-generalizable crossband channel prediction through a physics-guided deep unfolding design. When deployed directly from simulation to real-world measurements without retraining, GUIDE attains near-ground-truth beamforming gain at real-time inference speed and substantially outperforms all baselines. These findings suggest that GUIDE can serve as a practical and deployable channel-inference component for real-time AI-RAN.
5
Acknowledgments
The research work described in this paper was conducted in the JC STEM Lab of Advanced Wireless Networks for Mission-Critical Automation and Intelligence funded by The Hong Kong Jockey Club Charities Trust. The work of Ruiqi Kong is supported in part by Hong Kong ITC under Project PiH/277/25GS. The work of He Chen is supported in part by Hong Kong RGC under Project 24210524.
References [1] Arjun Bakshi, Yifan Mao, Kannan Srinivasan, and Srinivasan Parthasarathy. 2019. Fast and Efficient Cross Band Channel Prediction Using Machine Learning. In The 25th Annual International Conference on Mobile Computing and Networking (Los Cabos, Mexico) (MobiCom ’19). Association for Computing Machinery, New York, NY, USA, Article 37, 16 pages. [2] Avishek Banerjee, Xingya Zhao, Vishnu Chhabra, Kannan Srinivasan, and Srinivasan Parthasarathy. 2024. HORCRUX: Accurate Cross Band Channel Prediction. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking (Washington D.C., DC, USA) (ACM MobiCom ’24). Association for Computing Machinery, New York, NY, USA, 1–15. [3] Zikun Liu, Gagandeep Singh, Chenren Xu, and Deepak Vasisht. 2021. FIRE: enabling reciprocity for FDD MIMO systems. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking (New Orleans, Louisiana) (MobiCom ’21). Association for Computing Machinery, New York, NY, USA, 628–641. [4] Deepak Vasisht, Swarun Kumar, Hariharan Rahul, and Dina Katabi. 2016. Eliminating Channel Feedback in Next-Generation Cellular Networks. In Proceedings of the 2016 ACM SIGCOMM Conference (Florianopolis, Brazil) (SIGCOMM ’16). Association for Computing Machinery, New York, NY, USA, 398–411.