Electromagnetic Inverse Scattering from a Single Transmitter
Yizhe Cheng, Chunxun Tian, Haoru Wang, Wentao Zhu, Xiaoxuan Ma, Yizhou Wang
TL;DR
This paper introduces a novel data-driven framework for electromagnetic inverse scattering that achieves high-quality reconstructions from a single transmitter, overcoming traditional limitations of data scarcity and nonlinearity.
Contribution
The work presents the first successful deep learning-based method for high-quality electromagnetic inverse scattering from only one transmitter, enhancing robustness and practicality.
Findings
Outperforms existing methods in accuracy and robustness
Achieves high-quality reconstructions with a single transmitter
Demonstrates practical applicability in electromagnetic imaging
Abstract
Electromagnetic Inverse Scattering Problems (EISP) seek to reconstruct relative permittivity from scattered fields and are fundamental to applications like medical imaging. This inverse process is inherently ill-posed and highly nonlinear, making it particularly challenging, especially under sparse transmitter setups, e.g., with only one transmitter. While recent machine learning-based approaches have shown promising results, they often rely on time-consuming, case-specific optimization and perform poorly under sparse transmitter setups. To address these limitations, we revisit EISP from a data-driven perspective. The scarcity of transmitters leads to an insufficient amount of measured data, which fails to capture adequate physical information for stable inversion. Accordingly, we propose a fully end-to-end and data-driven framework that predicts the relative permittivity of scatterers…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. **Novel Problem Formulation:** This paper's primary contribution lies in its re-framing of the EISP problem. Instead of relying on traditional physics-based initializers (like BP) or slow, case-specific optimization, the authors propose a fully end-to-end, data-driven framework. This feed-forward model directly learns the mapping from scattered fields to permittivity by leveraging data distribution priors, which represents a clear paradigm shift from conventional approaches. 2. **Comprehensi
1. **Unclear Methodological Innovation:** The methodology section is brief, making it difficult to identify the core technical contributions. From my reading, the approach appears to be a straightforward application of NeRF-like implicit neural representations to the electromagnetic imaging domain. Similar methods already exist in the literature (e.g., NeRF2: Neural Radio-Frequency Radiance Fields), which casts doubt on the novelty of this work. The authors need to clearly articulate what distin
+ This work shows superior performance over all tested baselines (BP, SOM, PGAN, Physics-Net, Img-Interiors) on synthetic (MNIST, Circular) and real (Institut Fresnel) datasets. + It also shows high-quality reconstructions with a single transmitter in the result section, under specific experimental settings.
The reviewer has a few major concerns regarding the weakness: 1. The main content of this work is applying existing simple ML models (MLP) to solve electromagnetic inverse problems, rather than developing new ML algorithms or theory. Therefore, this work is more suitable for publication in the field of computaitonal electromagnetics rather than in machine learning. 2. This paper claims that it address the ill-posedness issue of electromagnetic inverse problem. This contribution is overclaimed. W
S1: The paper provides a clear formulation of the electromagnetic inverse scattering problem and presents the methodology in an accessible manner. The problem setup, including the forward scattering model and the inverse reconstruction task, is well-explained for readers unfamiliar with this domain. S2: The authors conduct experiments across multiple baseline methods. The visual comparisons demonstrate qualitative improvements over existing approaches in several scenarios. S3: The paper extend
W1: The core technical contribution of this paper is remarkably simple and lacks novelty. An MLP that takes scattered field measurements E^s and spatial coordinates x as input to predict permittivity, combined with standard positional encoding and a basic MSE loss on permittivity. The idea of training neural networks in a supervised, end-to-end fashion to learn direct mappings from input to desired outputs is a standard paradigm in machine learning and has been widely applied across various doma
(1)This approach addresses the challenging yet practical single-transmitter inverse scattering problem. The proposed end-to-end multilayer perceptron model replaces complex iterative solvers, significantly reducing inference time. (2)Covers diverse datasets — Circular, MNIST, 3D MNIST, 3D ShapeNet, and real Fresnel measurements — with multiple noise levels and training data sizes, showing reasonable robustness. (3)The paper is well-organized, and the writing is clear. The framework is clearly
（1） The paper mentions joint training on multiple datasets, but it’s unclear whether baselines were trained under the same setup or reused from prior papers. （2）While the paper emphasizes inference efficiency (e.g., much faster than Img-Interiors), it provides no quantitative analysis of the training cost, such as training time, GPU resources, dataset size, or convergence behavior. Since the proposed approach relies entirely on supervised learning with a large number of simulated examples, the
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsElectromagnetic Scattering and Analysis · Antenna Design and Optimization · Electromagnetic Compatibility and Measurements