AntennaFlow: A Generative Flow Model for Offset Correction in Phaseless Antenna Testing

Yongzhi Li; Chongting Shen; Menglin Chen; Xun Jiang; Zhengpeng Wang

cs.AISep 15, 2026

AntennaFlow: A Generative Flow Model for Offset Correction in Phaseless Antenna Testing

Authors: Yongzhi Li, Chongting Shen, Menglin Chen, Xun Jiang, Zhengpeng Wang

Organizations: College of Computing and Data Science Nanyang Technological University Singapore · Beihang Sino-French Engineer School Beihang University Beijing, China · School of Computer Science and Engineering University of Electronic Science and Technology of China Chengdu, China · Electronic Information Engineering Beihang University Beijing, China

Abstract

Near-field to far-field transformation is central to large-aperture antenna testing, yet two coupled challenges remain: costly phase acquisition at millimeter-wave bands and violations of the centering assumption under offset mounting. Existing methods address these issues separately, requiring either dense full-field data or offset vectors. We tackle both jointly by exploiting a key observation: amplitude fields under different offsets are coordinate-transformed views of the same near field. The challenge is to recover the center-aligned field from offset amplitudes without a phase or offset vector. We propose AntennaFlow, a three-stage framework: a contrastively learned encoder that maps offset views to an offset-invariant embedding, a deterministic flow-matching transport that maps offset amplitudes to center-aligned ones, and the Simplified Extrapolation Technique, whose Green-function Taylor expansion is valid only for centered fields. Experiments show that AntennaFlow enables fast, phaseless, offset-vector-free NF--FF reconstruction from sparse amplitude-only measurements, consistently outperforming existing baselines while preserving physical consistency.

Figures & tables

Fig. 1: Comparison of far-field reconstruction under offset mounting: prior methods suffer from offset sensitivity or reliance on offset vectors, while AntennaFlow achieves superior reconstruction quality.

Capabilities	Method
Capabilities	ET	SET	SWE	TSWE	SRM	Ours
Sparse Sampling	✗	✓	✗	✗	✗	✓
Fast Inference	✗	✓	✗	✗	✗	✓
Phaseless	✓	✓	✗	✗	✓	✓
Offset Correction	✗	✗	✗	✓	✓	✓
Offset-Vector Free	–	–	–	✗	✗	✓

TABLE I: Comparison of the previous methods and the proposed method

Fig. 2: The AntennaFlow pipeline for NF-FF transformation. (1) Data Collection: Near-field data are sampled from different antenna configurations. (2) A feature encoder is trained using a supervised contrastive learning framework to learn discriminative representations. (3) Flow Matching Transport: A U-Net-based architecture is used to learn the flow-matching velocity field for center near-field generation. (4) Far-Field Reconstruction via SET. Note that SET requires measurements from multiple distinct surfaces; for simplicity, only one surface is illustrated in the figure.

Fig. 3: Comparison of antenna far-field reconstruction using different phaseless reconstruction methods. The red solid curves denote the reconstructed far-field patterns, while the black dashed curves represent the theoretical far-field reference. The cyan curves indicate the corresponding ESS.

Fig. 4: Visual comparison of reconstructed far-field amplitude patterns under offset conditions. From left to right: SET, TSWE, SRM, AntennaFlow, and the ground truth. AntennaFlow demonstrates the highest fidelity in restoring detailed pattern textures compared to the ground truth.

Fig. 5: Reconstruction from offset near-field measurements. Different offset inputs (top) are mapped to consistent center-aligned fields (bottom).

Dataset	Metric	SET	TSWE	SRM	Ours
MATLAB	RMS ESS ( $\downarrow$ )	-18.97	-28.37	-34.44	-36.90
MATLAB	Peak ESS ( $\downarrow$ )	-6.91	-18.48	-24.32	-27.06
FEKO	RMS ESS ( $\downarrow$ )	-18.28	-25.25	-29.70	-33.81
FEKO	Peak ESS ( $\downarrow$ )	-8.42	-15.34	-18.33	-25.58
-	Inference Time	4s	10 min	30 min	9s

TABLE II: Quantitative Comparison of Far-Field Reconstruction Accuracy (RMS ESS and Peak ESS in dB)

CL Use	NF RMS	NF Peak	FF RMS	FF Peak
w/o CL	-25.74	-19.18	-24.43	-16.46
w/ CL	-34.39	-26.40	-33.81	-25.58

TABLE III: Ablation Study on the Effect of Contrastive Pre-training

Fig. 6: t-SNE projection of the encoder embeddings of $768$ amplitude samples drawn from $24$ distinct antenna configurations under random offsets. Colors denote antenna identity. Samples of the same antenna fall into tight clusters, while clusters of distinct antennas are clearly separated.

Perturbation	NF RMS	NF Peak	FF RMS	FF Peak
Default	$-35.39$	$-26.40$	$-33.81$	$-25.58$
Horn Probe	$-33.94$	$-25.89$	$-32.67$	$-24.50$
Dipole Probe	$-34.28$	$-24.48$	$-32.91$	$-24.81$
Additive noise	$-33.49$	$-25.18$	$-31.07$	$-23.67$

TABLE IV: Sim-to-Real Robustness Evaluation.(RMS ESS AND PEAK ESS IN DB)

Deeply unrolled alternating projections for phase retrieval
Oct 12, 2026Jiayi Yao, Vasileios Charisopoulos, Rebecca WillettImage Inverse Problems Deep Learning Optimization
Electrical & Computer Engineering, University of Washington · NSF-Simons National Institute for Theory and Mathematics in Biology · Department of Computer Science, University of Chicago +3
Physics-Augmented Graph Transformers for Patch-Antenna Forward and Inverse Design
Oct 4, 2026Avi Epstein, Snir Nehemia, Haim Suchowski +1Physics-Informed GNNs Graph Transformer
Tel Aviv University
AirPASS: Over-the-Air Federated Learning via Pinching Antenna Systems
Jul 7, 2026Seyed Mohammad Azimi-Abarghouyi, Christopher G. BrintonFederated Learning Aggregation Wireless Communications
Department of Electrical Engineering, Chalmers University of Technology, Gothenburg, Sweden · School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN USA

Oct 12, 2026J/K move · Enter open · S save

Deeply unrolled alternating projections for phase retrieval

Jiayi Yao, Vasileios Charisopoulos, Rebecca Willett

Electrical & Computer Engineering, University of Washington · NSF-Simons National Institute for Theory and Mathematics in Biology · Department of Computer Science, University of Chicago+3

Image Inverse Problems Deep Learning Optimization

Open paper AskPDF arXiv

Deep unrolling methods for inverse problems augment traditional reconstruction algorithms with learned priors to improve their performance. These methods have been developed under the implicit assumption that the geometry of the measurement operator does not change between training and deployment; when this is no longer the case, the quality of obtained reconstructions can degrade significantly. In this paper, we argue that this is not an inherent limitation of the deep unrolling paradigm and can be mitigated by a careful choice of the underlying iterative algorithm. We propose a deep unrolling framework for the phase retrieval problem that augments the alternating projections method with a learned denoising step. Our method obtains high-quality reconstructions without the need for retraining in settings where conventional model-based alternatives struggle, such as when the dynamic range of the measurement operator changes significantly at test time.

Abstract

Explore similar work

Deeply unrolled alternating projections for phase retrieval

Physics-Augmented Graph Transformers for Patch-Antenna Forward and Inverse Design

AirPASS: Over-the-Air Federated Learning via Pinching Antenna Systems