DeformRF - Data-driven Beamforming and Direction Finding with Deformable Antenna Arrays

ACM MobiSys'26

Abstract

Low-frequency antenna arrays enable obstacle-penetrating communications, but their large physical footprint—a 4×4 array at 150 MHz requires 16 m²—prevents practical deployment. DeformRF enables portable, deformable arrays that compress to backpack-size yet maintain beamforming performance when deployed on flexible substrates. Our key insight is that data-driven methods can predict complex electromagnetic behavior under deformation without real-time simulations. DeformRF combines: (1) a 260,000-sample synthetic dataset mapping deformations to EM characteristics, (2) physics-inspired ML models achieving >94% prediction accuracy on average, and (3) smartphone-based 3D reconstruction requiring no infrastructure. In real-world experiments, DeformRF maintains beamforming gains within 1 dB of optimal despite severe deformation, while baselines degrade by 4–10 dB. For emergency response scenarios, our 4×4 flexible array achieves ±5° direction-finding accuracy when tracking signals through buildings.

Deform2EM dataset and model training
DeformRF Deployment
DeformRF system overview

The Challenge

Conventional antenna array design assumes rigid substrates with precisely controlled geometry. Many real-world deployments, however, involve surfaces whose shapes are dictated by external constraints—building facades, vehicle bodies, and curved architectural surfaces make mounting rigid, high-gain antennas physically impractical. This problem is especially severe at lower frequencies (≤ 2.4 GHz), where obstacle-penetrating communications demand large apertures.

A fabric-based 4×4 array can be rolled into a 30 cm cylinder to fit a backpack, then rapidly unfolded and draped over any available surface. Yet this flexibility introduces a fundamental challenge: deformations alter electromagnetic (EM) behavior in complex, non-linear ways—simultaneously changing radiation patterns, impedance, and phase.

A flexible antenna array can be compactly stored and expanded during operation for beamforming and direction estimation. However, its EM behavior varies due to changing surface curvatures.

Key Challenges

Enabling effective wireless operation with deformable antenna arrays presents two significant challenges:

  1. Infinite deformation states — deformable substrates can assume virtually infinite configurations, making it impractical to pre-compute EM behaviors for all possibilities.
  2. Dynamic changes — deformations occur continuously during normal use as the substrate bends, folds, and twists, rendering one-time calibration methods ineffective.

Conventional EM simulations (e.g., Method of Moments) are computationally prohibitive—simulating a single deformed antenna can take several hours on a standard PC.

The DeformRF Approach

DeformRF is a lightweight end-to-end system with three core components:

1. Deform2EM Dataset

A comprehensive synthetic dataset of ~260,000 unique deformed antenna configurations and their EM characteristics, generated using Blender physics simulations and full-wave EM solvers. The dataset spans frequencies from 150 MHz to 850 MHz and covers both simple dipoles and complex Yagi–Uda arrays.

Deform2EM library curves are extracted from computer-graphic animations of DeformRF in Blender, complemented by control equations, and synthesized across 8 frequencies. Each curve (dipole and Yagi) is stored as a point cloud with four simulated EM parameters. The dataset contains more than 260,000 curves.

2. Physics-Inspired ML Model

A hybrid CNN–Vision Transformer architecture that predicts the complete EM behavior of a deformed antenna directly from its 3D geometry. Motivated by the computational structure of Method of Moments solvers:

  • CNN branch (ResNet-Lite, 6 residual blocks) captures local geometric features.
  • ViT branch (depth 8, 6 attention heads) captures global electromagnetic dependencies.
  • Four parallel MLP heads predict radiation pattern, phase, polarization, and reflection coefficient (Γ).
The CNN+Transformer hybrid model for DeformRF, predicting four EM parameters from pixelized antenna geometry.

3. Smartphone-Based 3D Reconstruction

DeformRF uses commodity smartphone cameras for one-time 3D capture of the antenna array, replacing expensive optical infrastructure. The pipeline:

  1. Spatial segmentation — isolates the antenna array from the background using tri-section clustering.
  2. RF element identification — locates antenna elements via RGB color anchoring.
  3. Curvature extraction — applies k-means clustering, hierarchical merging, and spline interpolation to recover each element’s precise 3D shape.
DeformRF curve reconstruction: We use a cell-phone to obtain the raw 3D image and perform background and floor removal, followed by RGB color anchoring and clustering, and finally centroid extraction and curve smoothening.

Key Evaluation Results


1. Beamforming Performance

Implemented 2×2 array: At low fold all methods exceed 12 dB, as the array is nearly flat. As deformation increases, DeformRF maintains consistent performance — outperforming DeformRF-PCA, Geometry, and Assume-Flat by >3.1 dB, 3 dB, and 4 dB respectively (a 3 dB gain effectively doubles array output). Prediction error stays below 1 dB across all fold states, while baselines reach 3–6 dB error under mid/high fold.

Simulated 5×6 array: DeformRF sustains ~30 dB gain across all 810 deformation configurations. DeformRF-PCA and Geometry drop to ~25 dB at mid/high fold; Assume-Flat falls below 20 dB. Prediction error holds at ~2 dB, while baselines exceed 5 dB as deformation increases.

(a) Impl: Measured gain
(b) Impl: Prediction error
(c) Sim: Measured gain
(d) Sim: Prediction error
(e) Sim: Beam-pattern
Performance of Beamforming: (a) and (b) show the mean gain and mean error for three folds for the implemented 2×2 array, while (c) and (d) show the mean gain and mean error for the simulated 5×6 array. (e) Beampattern of an instance of the 5×6 antenna array compared with the PCA baseline.


2. Direction-of-Arrival Estimation

Indoor 4×4 array (850 MHz): DeformRF achieves ±5° DOA accuracy, outperforming baselines by more than 3×.

Outdoor-to-indoor emergency scenario: 4° DOA error — matching commercial rigid arrays despite deformation, and nearly halving the best baseline.

(a) Deform2EM DOA result and estimated room coverage.
(b) DOA est. error across five tested angles (x-axis).
(a): Left — Average DOA results of DeformRF compared to the flat assumption. Right — Estimated room coverage for DeformRF and the assume flat. (b): DOA accuracy at each test angle.

Key Contributions

  • Deform2EM Dataset — First large-scale synthetic dataset (260k samples) mapping physical antenna deformations to full EM characteristics across 150–850 MHz.
  • Physics-Inspired ML — Hybrid CNN–ViT model achieving >94% EM prediction accuracy, running in real-time on commodity hardware.
  • Infrastructure-Free Sensing — Smartphone-only 3D reconstruction pipeline achieving sub-wavelength deformation accuracy without LiDAR or optical base stations.
  • End-to-End System — Integrated beamforming and DOA estimation that maintains performance within 1 dB of optimal under severe deformation.