Research map / Probabilistic and scientific modeling / Dynamical systems
NEAR: Neural Electromagnetic Array Response
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NEAR, an untrained implicit neural representation with a physics-informed block-Hankel low-rank regularizer that predicts MIMO radar virtual-array responses at unseen locations from sparse measurements.
Problem
Angular super-resolution in multi-antenna (MIMO) radar with few physical antennas and sparse sampling; supervised learning generalizes poorly to unseen environments and existing matrix-completion and NeRF-style methods underperform.
Approach
Train a randomly initialized INR (NeRF-style sinusoidal positional encoding MLP) per scene on sparse measurements, adding a implicit regularizer that enforces low rank of block Hankel matrices of the predicted response field via learnable least-squares coefficients, grounded in a tight characterization of INR-representable harmonics.
Key results
- NEAR achieves NRMSE 0.0495 at 20 dB with 8x8 sampling vs 0.1889 for EMaC and about 1.05 for SIREN/NEAR w/o R.
- In real-world angular resolution at 2 m, NEAR matches the full-array benchmark at 5.7248° while EMaC needs 8.5783° and NeRF2 8.5783°.
- Localization mean absolute error for K=1 is 0.0744 for NEAR vs 0.0827 for the full-array benchmark and 0.1062 for EMaC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NEAR instead of EMaC for sparse response recovery | Simulation response recovery · 20 dB, 8×8 sampling, 2 targets | NRMSE: 0.1889 -> 0.0495 | Helped |
| Use NEAR instead of EMaC for real-world angular resolution | Radar angular resolution · 4.5M | Smallest resolvable angular separation: 10.1592◦ -> 6.9941◦ | Helped |
| Use NEAR instead of EMaC for real-world localization | Radar localization · K=4 targets | Localization error: 0.1157 -> 0.0718 | Helped |
| Use NEAR instead of EMaC for computation time | Radar response recovery · Computation time | Averaged time cost (s): 1226.15 -> 550.83 | Helped |
| Add physics-informed regularization (ablation) | NEAR · Simulation response recovery, 20 dB, 8×8 sampling, 2 targets | NRMSE: 1.0504 -> 0.0495 | Helped |
| Add physics-informed regularization (ablation) | NEAR · Simulation response recovery, 20 dB, 6×6 sampling, 2 targets | NRMSE: 1.0689 -> 0.1884 | Helped |
| Add physics-informed regularization (ablation) | NEAR · Simulation response recovery, 20 dB, 8×8 sampling, 3 targets | NRMSE: 1.0308 -> 0.0860 | Helped |
Surprising findings
- NEAR achieves lower response error than the noisy 20x20 full virtual array benchmark, which the authors attribute to the regularizer's denoising effect.
- Plain INR without the regularizer (SIREN, NEAR w/o R) performs near chance (NRMSE about 1.0), showing that data fitting alone fails with sparse measurements.
Limitations
- Computational cost remains high for real-time embedded deployment (authors flag as future work).
- Regularizer needs only an upper bound on target count, but performance degrades with more targets in simulation (NRMSE rises from 0.038 to 0.086 from 1 to 3 targets).
- Real-world evaluation uses a single commercial radar platform and a limited set of corner-reflector scenes.
Experimental setup
Datasets: simulated array responses with 2 targets (noisy) real-world measurements from IMAGEVK-74 MIMO radar (20x20 virtual array)
Benchmarks: simulated virtual array response recovery (NRMSE) angular resolution probability DOA estimation error real-world angular resolution real-world target localization
Metrics: NRMSE resolution probability DOA estimation error (deg) angular resolution (deg) mean absolute localization error running time
Baselines: EMaC SIREN NeRF2 NEAR w/o regularization full virtual array benchmark
Builds on: NeRF SIREN Fourier feature networks enhanced matrix completion NeRF2
Closest papers by meaning
- RaUF: Learning the Spatial Uncertainty Field of Radar (CVPR 2026 · similarity 0.84)
- Physics Unrolled Neural Operator for Wireless Field Modeling (NeurIPS 2026 · similarity 0.82)
- Implicit Neural Representations and the Algebra of Complex Wavelets (ICLR 2024 · similarity 0.82)
- Hierarchical Neural Operator Transformer with Learnable Frequency-aware Loss Prior for Arbitrary-scale Super-resolution (ICML 2024 · similarity 0.82)
- DART: Implicit Doppler Tomography for Radar Novel View Synthesis (CVPR 2024 · similarity 0.82)
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.