Real measurements are too limited to train an accurate indoor pathloss predictor that works across a wide range of buildings. Large-scale synthetic pretraining offers a way forward, but real building layouts are also scarce, and accurate ray-tracing simulators are slow, expensive, and often proprietary. We address both problems by introducing a generator that automatically creates diverse indoor layouts and a fast, open-source approximate ray tracer that computes radio maps for them. Using this pipeline, we generate 128M synthetic samples, pretrain ResNet-101 and Swin-T encoder-decoders, and fine-tune them on high-quality benchmark data. Despite simulator mismatch, our ResNet-101 model reduces error by 15% on average relative to the best known methods across five indoor radio-map prediction tasks and improves transfer on a small real-measurement dataset. Both architectures show predictable data-scaling trends: power laws fitted on 2M-16M samples predict the mean normalized downstream RMSE at 32M-128M with 3.33% mean absolute percentage error for ResNet-101 and 7.10% for Swin-T. These results show that pretraining on large amounts of cheap simulated data can improve indoor radio-map prediction, even when real measurements and high-quality simulations are limited.
Figures & tables
Task
Setting
Train
Val
Public
ICASSP T1
Unseen buildings; 868 MHz, isotropic
1,000
250
50
ICASSP T2
Unseen buildings/frequencies; three training frequencies, test includes 2,400 MHz
3,000
750
100
ICASSP T3
Unseen buildings/frequencies/antennas; five training and two test patterns
22,200
5,550
200
MLSP 0.02
Sparse pathloss at 0.02% of pixels
3,000
750
50
MLSP 0.5
Sparse pathloss at 0.5% of pixels
3,000
750
50
Table 1: Downstream tasks and map counts. ICASSP and MLSP use high-fidelity simulated targets ( Bakirtzis et al., 2025 ; Yapar et al., 2025 ) ; Public is the Kaggle evaluation subset.
Figure 1: Representative synthetic (top) and ICASSP benchmark (bottom) layouts, antenna-pattern channels, and pathloss targets. The examples have different geometry and use separate colour scales; panels preserve aspect ratio but not a common physical scale. ICASSP targets are high-fidelity simulations. The full synthetic input-channel illustration is in Appendix A .
Encoder
Peak LR
Batch
FLOPs/sample
Train loss
Val RMSE
ResNet34
3×10−3
355
4.829
0.03356
11.837
Swin-T p4-w4
10−3
321
5.340
0.03441
11.913
Swin-T p4-w5
10−3
294
5.831
0.03367
11.894
Swin-T p4-w7
10−3
272
6.303
0.03375
11.885
ResNet101
3×10−3
181
9.471
0.03222
11.774
ConvNeXt-T
3×10−3
142
12.07
0.03327
11.856
Table 2: Fixed-compute search: 350k steps and 6×1018 FLOPs. Each row minimizes final training loss over the hyperparameter sweep; all selected runs use weight decay 10−3 . FLOPs/sample are in units of 1010 ; validation RMSE is in dB. For Swin-T, p and w denote patch and attention-window sizes.
Figure 2: Downstream scaling at all seven first-epoch budgets for ResNet101 and Swin-T. The first five panels show each task’s public-leaderboard RMSE; vertical ranges differ by task. The final panel averages task RMSE normalized by each backbone’s own 2M result. Lines connect observations; the dotted divider marks the end of the fitting range (2M–16M). Swin-T’s lower normalized RMSE mainly reflects its higher 2M errors and larger relative improvement, while mean RMSEs at 128M are fairly similar for both backbones.
Task
Reference entry
Scratch
Ref.
Best
Scale
Reduction
ICASSP T1
C. T. Cisse
5.484
4.908
4.311
128M
12.17%
ICASSP T2
takktak
9.326
10.307
7.435
64M
27.86%
ICASSP T3
takktak
13.302
12.018
9.703
32M
19.27%
MLSP 0.02
George Hoa
7.736
4.351
3.993
128M
8.23%
MLSP 0.5
Leyi Zou
5.872
3.731
3.203
128M
14.16%
Table 3: Public-leaderboard RMSE (dB; lower is better). Best denotes the lowest score over first-epoch pretraining scales after internal-validation epoch selection. Reduction is relative to the paper’s leaderboard reference, not the full hidden challenge test set.
Figure 6
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Representative synthetic inputs and dense pathloss target. Top: reflectance, transmittance, and radiation pattern. Bottom: floor plan and target. Pretraining and fine-tuning use the same channel semantics.
Backbone
2M
4M
8M
16M
32M
64M
128M
ResNet101
7.91
7.64
6.62
6.58
6.58
6.18
6.45
Swin-T
9.57
8.50
7.89
7.31
6.55
6.78
6.95
ResNet101
1.000
.959
.877
.869
.873
.805
.833
Swin-T
1.000
.850
.793
.747
.658
.677
.694
Appendix
Table 4: Five-task means at every first-epoch scale. Upper rows: raw RMSE (dB). Lower rows: mean of per-task 2M-normalized RMSE. Best raw means occur before 128M.
Evaluation
0 reflections
2 reflections
5 reflections
Direct simulator (ICASSP median)
12.61
11.11
10.91
Frozen transfer (five-task mean)
10.50
9.02
8.99
Full fine-tuning (five-task mean)
6.79
6.60
6.58
Appendix
Table 5: Reflection-depth comparison (RMSE, dB). Direct simulation reports median per-map error; transfer reports five-task mean error. Full fine-tuning substantially narrows the fidelity gap.
Task
Labeled maps
0 reflections
2 reflections
5 reflections
ICASSP T2
60
9.68
8.11
7.81
120
9.53
8.12
7.85
300
7.53
7.08
7.16
720
7.31
6.95
6.92
1,500
6.89
6.76
6.89
MLSP 0.02
60
8.21
7.58
7.25
Appendix
Table 6: Scarce-label transfer from 16M checkpoints (RMSE, dB; mean of two seeds). Label counts 60, 120, 300, 720, and 1,500 correspond to 2%, 4%, 10%, 24%, and 50% of 3,000 maps, respectively.
Support
Input
2M
4M
8M
16M
0%
Full
10.366
9.711
10.264
10.776
0%
No R/T
15.776
18.882
20.669
21.615
0.02%
Full
8.722
8.104
8.647
9.238
0.02%
No R/T
9.232
9.157
9.430
10.044
0.5%
Full
5.101
4.623
4.539
4.453
0.5%
No R/T
5.160
4.687
4.587
4.530
Appendix
Table 7: Frozen MLSP RMSE (dB) with full inputs or no material channels (No R/T). At zero support, error rises through 32M without materials: additional pretraining does not guarantee transfer.
Exposure
RMSE (dB)
95% AP-cluster interval
pHolm vs. 8M
8M
5.117
[4.583, 5.642]
—
128M
5.323
[4.727, 5.908]
0.304
256M
5.234
[4.671, 5.779]
1.000
Appendix
Table 8: Combined real-data models and AP-group uncertainty. The later checkpoints are not distinguishable from 8M after multiple-comparison correction; nonsignificance does not establish equivalence.
Figure 5: Real RSSI regression across synthetic-pretraining scales. The neural network is frozen; each regressor is trained on the training AP locations in its fold. The best numerical combined RMSE occurs at 8M, but comparisons with 128M and 256M do not resolve their ordering.
Figure 6: Out-of-fold RSSI predictions for the 8M combined ridge model. The diagonal indicates perfect agreement; deviations show remaining prediction error. This is learned target calibration with frozen neural features, not end-to-end transfer.
Figure 7: Free-space distance scaling probe. Across 1–100 m and 900, 2400, and 5800 MHz, annulus-median pretrained-model predictions track analytic FSPL with aggregate RMSE 0.80 dB, 1.10% MAPE, and R2=0.996 .
Figure 8: Open-field rotational isotropy probe. After subtracting the radial median from each annulus, the pretrained model has small angular residuals: median absolute angular residual is 0.037 dB and the 95th percentile is 0.15 dB.
Figure 9: Free-space frequency scaling at fixed distances. Annulus-median pretrained-model predictions track analytic FSPL with 0.45% MAPE and R2=0.999 .
Figure 10: Finite 12 dB wall shadowing probe. The predicted blocked-minus-unblocked contrast is 11.57 dB for the pretrained model and 8.71 dB for the Task 3 baseline.
Figure 11: Single-wall attenuation calibration across 3–20 dB wall losses. The pretrained model has mean median absolute error 0.41 dB across the sweep and mean R2=0.987 .
Figure 12: Empty-room 2.4 GHz diagnostic scene compared with analytic FSPL.
Figure 13: Single 10 dB vertical wall diagnostic scene compared with FSPL plus geometric behind-wall loss.
Figure 14: Empty-room FSPL with four exact sparse conditioning supports.
Figure 15: D4 scene-isometry consistency probe. Panels show prediction differences after aligning transformed predictions back to the original frame.
Figure 16: Source-receiver reciprocity probe over 231 endpoint swaps in one representative layout.
Radio maps are essential for wireless decision-making tasks such as access-point placement, coverage planning, and localization, but their fine spatial details are governed by complex propagation effects and are costly to simulate accurately. Machine learning offers a path to high-fidelity radio-map prediction without running expensive high-fidelity simulations for every scene. However, generating high-quality training labels at scale is also difficult: the affordable labels come from finite-ray simulations, which are richer than low-fidelity inputs but carry residual Monte Carlo noise. We address this challenge with Physics-Unrolled Hybrid Neural Operator (PU-HNO), a three-stage cascade that predicts high-fidelity indoor radio maps from low-fidelity ray-tracing outputs and scene priors by progressively capturing reflection, diffraction, and scattering effects, rather than treating radio maps as generic images. We prove that, under conditionally unbiased label noise, the model can learn stable propagation structure and outperform its own training labels. Experiments across diverse floorplans show that PU-HNO outperforms image-to-image baselines, wireless learning models, and monolithic neural operators across both image-quality and wireless deployment metrics.
Rafid Umayer Murshed, Saif Ur Rahman, Mingyue Tang +1
Department of Computer Science University of Illinois Urbana-Champaign Urbana, IL 61801, USA
Radio maps describe how wireless signals propagate across space and are essential for wireless communication, sensing, and network planning. However, constructing accurate radio maps traditionally requires either dense measurements or computationally expensive physical simulations, which limits scalability and real-time deployment. Recent advances in generative artificial intelligence offer a promising alternative, but existing approaches lack fine-grained control and physical consistency when applied to real-world wireless environments. Here we present ControlRadio, a controllable generative framework that produces radio maps from natural-language descriptions and environmental layouts, including building structures and transmitter locations. Joint semantic and spatial conditioning enables interpretable, propagation-plausible generation, while a controlled latent prior and layout-aware conditioning improve stability and structural consistency. Extensive experiments demonstrate that ControlRadio achieves state-of-the-art accuracy and strong generalization across diverse urban scenarios, while reducing computation time by more than four orders of magnitude compared with conventional simulation-based methods. Such results suggest a new paradigm for scalable and controllable wireless environment modeling, with broad implications for next-generation communication systems and data-driven radio sensing.
Kangjun Liu, Xiying Pan, Shuhang Zhang +3
Pengcheng Laboratory, China · South China University of Technology and the Pengcheng Laboratory, China · Peking University, China +2
Accurate path loss prediction is a critical component of wireless network planning. Current path loss prediction methods typically struggle to balance the trade-off between accuracy and computational efficiency. This paper proposes the Efficient Attention Radio Map Estimation Network (EA-RMENet) which is an image data-driven, deep learning (DL) model designed for radio map estimation (RME). EA-RMENet uses a U-Net framework with an EfficientNetB5 encoder, Attention Gated (AG) skip connections, and Atrous Spatial Pyramid Pooling (ASPP). The EfficientNet encoder uses compound scaling to balance accuracy and efficiency. AG skip connections suppress irrelevant features, and the ASPP captures a multi-scale context. The model has a test prediction RMSE of 0.0334 on the RadioMapSeer3D dataset with an inference time of 0.022 seconds/sample. In the ICASSP 2023 Radio Map Prediction Challenge, the model ranks third with a competitive RMSE of 0.0406 this highlights the models potential for real-world RME.