RFWeave: Learning Wireless Propagation Across Scenes and Transceiver Configurations

Shurui Xu; Xiao Ye; Yijie Bian; Junjie Ai; Yuxuan Chen; Weicong Chen; Wankai Tang; Shi Jin

eess.SPOct 12, 2026

RFWeave: Learning Wireless Propagation Across Scenes and Transceiver Configurations

Authors: Shurui Xu, Xiao Ye, Yijie Bian, Junjie Ai, Yuxuan Chen, Weicong Chen, Wankai Tang, Shi Jin

Organizations: Southeast University · The Hong Kong University of Science and Technology

Abstract

Wireless propagation modeling is foundational to the evaluation and design of modern communication systems. To achieve accurate wireless propagation prediction, recent learning-based approaches leverage scene geometry and materials, demonstrating promising generalization across transceiver locations and unseen scenes. However, these approaches remain largely confined to a fixed configuration of array size, receiver orientation, and transmit beam. In this work, we introduce RFWeave, a learned wireless propagation model designed to generalize across scenes, transceiver locations, and transceiver configurations. To accommodate varying array structures, RFWeave incorporates variable-length transceiver tokens that represent antenna elements individually. Furthermore, to capture the geometry between transmit and receive antenna elements, RFWeave aggregates their pairwise spatial relations into propagation tokens. A configuration-conditioned decoder predicts the spatial spectrum and received power. RFWeave is trained on a large-scale simulated dataset spanning 10,000 indoor scenes and more than 5 million channel records under diverse transceiver configurations. Extensive evaluations show that RFWeave achieves state-of-the-art performance across all evaluated generalization settings, improving spatial spectrum structural similarity (SSIM) by up to 15.7% and reducing received power mean absolute error (MAE) by up to 35.4% over the strongest baselines on unseen scenes. Additional experiments on measured DICHASUS data and in beam management further demonstrate RFWeave's broader applicability to real-world channels and downstream wireless tasks.

Figures & tables

Figure 1: RFWeave. Given a material-annotated scene, queried TX/RX locations, and a transceiver configuration, RFWeave predicts the AoA spatial spectrum and received power using a single model. The same model supports arbitrary TX/RX location queries and generalizes across array sizes, receiver orientations, transmit beams, and their combinations on unseen scenes. The received power heatmap is obtained by querying RFWeave at multiple receiver locations.

Figure 2: Overview of RFWeave. RFWeave organizes scene and transceiver information into structured tokens and refines element-indexed propagation tokens with a scene-conditioned backbone. The prediction stage outputs received power and a complex covariance factor, from which the spatial covariance and AoA spatial spectrum are obtained.

Figure 3: Comparison on an unseen scene. We compare received power and spatial spectrum from ground truth, RFWeave, and Wi-GATr under different transceiver configurations. (a) uses transmit beam index 0, receiver orientation $180^{\circ}$ , and a $4\times 4$ UPA. Panels (b) – (d) vary one configuration factor from this reference: (b) changes the transmit beam index to 5, (c) changes the receiver orientation to $45^{\circ}$ , and (d) increases the array size to $5\times 5$ . Further analysis is provided in Sec. 4.1 .

Method	Per-scene Opt.	WiPTR						Wi3R
		Held-out ID Scenes			OOD Scenes			Held-out ID Scenes			OOD Scenes
		SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$
+ Proposed Decoder
PointNet++	None	0.7988	3.16	4.75	0.7907	3.26	4.51	0.7414	4.00	8.61	0.7073	4.43	10.91
PointNeXt	None	0.7993	3.16	4.34	0.7924	3.25	4.37	0.7421	3.96	8.60	0.7092	4.38	10.87
Wi-GATr	None	0.7996	3.13	3.92	0.7963	3.14	4.02	0.7371	4.03	8.62	0.5821	5.77	11.46

Table 1: Cross-scene propagation prediction on WiPTR and Wi3R. The upper block evaluates feed-forward prediction, while the lower block allows 100 target-scene samples for adaptation.

Method	Array Size			Receiver Orientation			Transmit Beam
Method	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$
Seen configurations, unseen scenes
PointNet++	0.7583	3.63	4.84	0.7848	3.35	4.35	0.7521	3.71	4.36
PointNeXt	0.7563	3.65	4.52	0.7909	3.28	4.09	0.7520	3.69	4.34
Wi-GATr	0.7633	3.56	3.93	0.7843	3.34	3.58	0.7526	3.68	3.60
Transformer	0.7411	3.84	4.71	0.7690	3.51	4.34	0.7297	3.93	4.48

Table 2: Single-factor configuration generalization on unseen WiPTR scenes. The upper and lower blocks evaluate seen and unseen configurations, respectively.

Method	Seen Configurations			Angular Interpolation			Array Size Extrapolation
Method	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$
Unseen scenes
PointNet++	0.6952	4.41	4.74	0.6540	4.89	6.24	0.5244	6.84	5.33
PointNeXt	0.6968	4.40	4.55	0.6585	4.87	6.22	0.5267	6.87	5.27
Wi-GATr	0.6866	4.48	3.53	0.6519	4.93	5.75	0.5396	6.75	4.93
Transformer	0.6712	4.69	4.58	0.6444	5.01	6.01	0.5250	6.98	5.48

Table 3: Joint configuration generalization on unseen WiPTR scenes under seen configurations, angular interpolation, and array size extrapolation.

Variant	SSIM $\uparrow$		MAE Spec $\downarrow$		MAE Pwr $\downarrow$
Variant	Value	$\Delta$ (%)	Value	$\Delta$ (%)	Value	$\Delta$ (%)
RFWeave-B (full)	0.8060	–	3.12	–	2.28	–
w/o material properties	0.7932	$-1.6$	3.31	$+6.1$	2.43	$+6.6$
w/o pairwise geometry	0.7756	$-3.8$	3.48	$+11.5$	2.67	$+17.1$
w/o configuration conditioning	0.7618	$-5.5$	3.69	$+18.3$	2.86	$+25.4$
Direct covariance supervision	0.7827	$-2.9$	3.41	$+9.3$	2.59	$+13.6$

Table 4: Ablation of the main RFWeave design choices.

\Delta

denotes the relative change (%) from the full model. Negative SSIM changes and positive MAE changes indicate degradation.

Training samples	RFWeave-B (FT)			NeRF 2			WRF-GS
Training samples	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$
100	0.7769	2.43	1.46	0.6689	2.91	2.64	0.6428	3.76	3.03
500	0.7820	2.28	1.38	0.7407	2.34	1.54	0.6433	3.53	2.76
1000	0.7836	2.28	1.35	0.7443	2.34	1.47	0.6662	3.41	2.53
5000	0.7921	2.27	1.31	0.7471	2.31	1.26	0.6710	3.20	2.46

Table 5: Adaptation to measured DICHASUS channels with different numbers of training samples.

Strategy	Top-1	Top-2	Top-3
Exhaustive sweep (GT)	14.64 / 14.86	–	–
Location-based steering	7.24 / 7.31	–	–
RFWeave-B (FT)	9.74 / 10.30	10.33 / 11.26	10.78 / 11.61

Table 6: Beam management with a

4\times 4

DFT codebook. Mean / median spectral efficiency (SE) in bps/Hz. Top-

K

retains the

K

highest-ranked beam pairs and reports the best.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Data type	Input parameterization	Tokenization	Type vector
Scene	${}^{G}\mathbf{p}$ , ${}^{G}\mathbf{n}$ , $\bm{\mu}$	1 token per final sampled center	$\mathbf{t}_{\mathrm{scene}}$
TX/RX elements	${}^{T}\mathbf{p}_{i}/\lambda$ , ${}^{R}\mathbf{p}_{j}/\lambda$	1 token per antenna element	$\mathbf{t}_{\mathrm{te}},\,\mathbf{t}_{\mathrm{re}}$
Array centers	${}^{G}\mathbf{c}^{\mathrm{t}}$ , ${}^{G}\mathbf{c}^{\mathrm{r}}$	1 token per array	$\mathbf{t}_{\mathrm{tc}},\,\mathbf{t}_{\mathrm{rc}}$
TX–RX link	${}^{G}\bm{\ell}$	1 token per link	$\mathbf{t}_{\mathrm{link}}$

Appendix

Table 7: Scene and transceiver tokenization. Token roles are distinguished by one-hot vectors.

Model	Hidden dim.	Heads	Prop. stages	Backbone (M)	Decoder (M)	Total (M)
RFWeave-T	128	8	6	3.62	1.27	4.89
RFWeave-S	192	8	6	7.38	2.70	10.08
RFWeave-B	256	8	8	16.63	5.04	21.67

Appendix

Table 8: Architecture and parameter counts of the RFWeave variants. Scene self-attention uses two blocks and the covariance factor rank is

r=16

for all variants.

Regime	Array size	Receiver orientation	Transmit beam
Training	6 seen sizes ( $3\times 3$ – $8\times 8$ )	8 seen orientations	16 seen directions
Single-factor: array size	4 unseen sizes ( $9\times 9$ – $12\times 12$ )	fixed	fixed
Single-factor: receiver orientation	fixed	16 unseen orientations	fixed
Single-factor: transmit beam	fixed	fixed	9 unseen directions
Joint: seen	6 seen sizes ( $3\times 3$ – $8\times 8$ )	8 seen orientations	16 seen directions
Joint: angular interpolation	6 seen sizes ( $3\times 3$ – $8\times 8$ )	8 seen orientations	9 unseen directions

Appendix

Table 9: Training and evaluation configurations for generalization experiments. Seen denotes values observed during training, unseen denotes held-out values used for generalization evaluation, and fixed denotes the reference value held constant in the corresponding single-factor experiment.

Steps	WiPTR Held-out ID Scenes			Wi3R OOD Scenes
Steps	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$
500	0.9398	1.498	0.962	0.8858	2.069	2.782
1000	0.9372	1.576	0.970	0.8828	2.113	2.115
2000	0.9407	1.556	1.045	0.8603	2.438	3.010

Appendix

Table 10: Fine-tuning sensitivity to the number of optimization steps. All settings use the same 100 target-scene samples and the fine-tuning configuration adopted in Table 1 .

Method	Params. (M)	$4\times 4$ GFLOPs	$8\times 8$ GFLOPs
PointNet++	5.19	21.05	67.49
PointNeXt	8.52	28.53	74.97
Wi-GATr	6.47	37.73	84.16
Transformer	8.22	21.50	67.93
RFWeave-T	4.89	10.36	28.99
RFWeave-S	10.08	16.80	47.90

Appendix

Table 11: Model size and total forward computation at

4\times 4

and

8\times 8

array sizes.

Figure 6: Received power MAE across array sizes. Models are trained on square UPAs with

N=3,\ldots,8

and additionally evaluated on unseen array sizes

N=9,\ldots,12

Figure 7: Computational scaling. Peak GPU memory is shown on the left and total forward GFLOPs on the right for square UPAs with varying side length

N

Figure 8: Accuracy–efficiency trade-off. Spatial spectrum SSIM is shown on the left and received power MAE on the right against total forward GFLOPs. Bubble size indicates the number of model parameters.

$\lambda_{\mathrm{rx}}$	$\lambda_{\mathrm{beam}}$	Unseen RX Orientation			Unseen TX Beam
$\lambda_{\mathrm{rx}}$	$\lambda_{\mathrm{beam}}$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$	SSIM $\uparrow$	MAE Spec $\downarrow$	MAE Pwr $\downarrow$
0	0	0.7128	3.809	2.692	0.7205	4.051	6.789
1	0	0.7485	3.525	2.771	0.7260	3.965	6.622
2	0	0.7471	3.513	2.785	0.7144	4.094	6.911
0	1	0.7138	3.876	3.052	0.7472	4.045	7.197
0	2	0.7171	3.812	2.879	0.7349	3.955	6.845

Appendix

Table 12: Sensitivity to the configuration variation loss weights.

Figure 9: Cross-scene prediction and target-scene adaptation. Qualitative comparison on an unseen scene. Each method predicts the received power map and the AoA spatial spectrum at the marked receiver location. RFWeave denotes the feed-forward prediction, while RFWeave (FT) is fine-tuned using 100 target-scene samples. NeRF 2 and WRF-GS are optimized independently for the target scene using the same samples. RFWeave better preserves the narrow high-power structure and separated spectrum lobes than the feed-forward baselines.

Figure 10: Transmit beam generalization. Feed-forward predictions on an unseen WiPTR scene as the transmit beam changes from index 0 to 5, while array size and receiver orientation remain fixed.

Figure 11: Receiver orientation generalization. Feed-forward predictions on an unseen WiPTR scene as the receiver orientation changes from the reference orientation to $45^{\circ}$ , while array size and transmit beam remain fixed.

Figure 12: Array size generalization. Feed-forward predictions on an unseen WiPTR scene as the square UPA changes from $4\times 4$ to $5\times 5$ , while receiver orientation and transmit beam remain fixed.

Radio Propagation Modeling

Radio Propagation Modeling predicts how electromagnetic signals travel through environments for wireless planning, channel estimation, and sensing.

See all 61 papers

MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation
Aug 5, 2026Blessed Guda, Kayley Sze, Carlee Joe-WongWireless Foundation Models Wireless Localization
Carnegie Mellon University Pittsburgh, USA
PropSplat: Map-Free RF Field Reconstruction via 3D Gaussian Propagation Splatting
May 8, 2026William Bjorndahl, Maninder Pal Singh, Farhad Nouri +13D Gaussian Splatting Wireless Communications
Department of Electrical and Computer Engineering, Southern Methodist University, Dallas, TX, USA · Department of Engineering Technology, University of Houston, Houston, TX, USA
RF-GPT: Teaching AI to See the Wireless World
Feb 16, 2026Hang Zou, Yu Tian, Bohao Wang +4Wireless Communications Language Modeling
Research Institute for Digital Future, Khalifa University, 127788 Abu Dhabi, UAE · College of Information Science and Electronic Engineering, Zhejiang University, 310027, Hangzhou, China · Université de Lorraine, CNRS, CRAN, F-54000 Nancy, France

Aug 5, 2026J/K move · Enter open · S save

MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation

Blessed Guda, Kayley Sze, Carlee Joe-Wong

Carnegie Mellon University Pittsburgh, USA

Wireless Foundation Models Wireless Localization Radio Propagation Modeling Channel Estimation

Open paper AskPDF arXiv

Recent advances in machine learning have enabled training of wireless foundation models, which aim to support tasks such as channel estimation, beam prediction, and localization based on wireless signals. Existing wireless foundation models typically pretrain on channel tensors using masked reconstruction over subcarriers, antennas, or time but ignore the physical characteristics of wireless propagation. In this work, we propose to instead use multipath propagation as the fundamental pretraining object. We present MultiPathFormer, an autoregressive foundation model that represents each transmitter-receiver link as an ordered sequence of continuous-valued path tokens and pretrains with next-path prediction. We introduce an Environmental RAG (retrieval-augmented generation) mechanism and a first-path codebook on top of the transformer backbone, leveraging environment knowledge to improve path statistics estimation like delay and power by up to 59%. MultiPathFormer pretrained on 27 environments transfers to unseen users and, after scenario-specific fine-tuning, outperforms training the corresponding models from scratch in new environments. Across downstream tasks, it outperforms SOTA channel-based foundation models, achieving 5.57 m mean localization error, 0.914 top-3 beam accuracy, 0.994 line-of-sight classification accuracy, and 0.561 channel estimation NMSE. These results show that path-level pretraining can learn reusable representations of wireless propagation.

Abstract

Radio Propagation Modeling

Explore similar work

MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation

PropSplat: Map-Free RF Field Reconstruction via 3D Gaussian Propagation Splatting

RF-GPT: Teaching AI to See the Wireless World