AI Solves Complex Electromagnetic Puzzles with Circular Neural Networks

In a groundbreaking study, researchers have harnessed the power of convolutional neural networks to tackle one of electromagnetism's toughest s: reconstructing hidden objects from scattered wave data. The paper, titled 'Inverse Electromagnetic Scattering for Doubly-Connected Cylinders using Convolutional Neural Networks,' introduces a novel divide-and-conquer approach that combines classification and regression to solve inverse problems in electromagnetic scattering. This addresses the ill-posed nature of such problems, where small errors in data can lead to large inaccuracies, by leveraging specially designed 1D multi-channel CNNs with circular padding to handle the periodicity of angular measurements. span fields like medical imaging and remote sensing, offering a non-invasive way to detect defects or characterize materials without direct contact.

Ology centers on a two-step process: first classifying the shape of an impedance cylinder—such as peanut, kite, or star-shaped—and then reconstructing its boundary curve and impedance function from far-field data. Using a circular-padding CNN architecture, the network processes real and imaginary components of electric and magnetic fields as separate input channels, preserving the 2π-periodicity of measurements taken around a unit circle. This design allows the model to learn cross-correlations between field components and capture multi-scale angular dependencies through layers with progressively increasing kernel sizes. Training data were generated by solving direct scattering problems via boundary integral equations, with datasets ranging from 30,000 to 120,000 samples, depending on complexity, and inputs including sparse (32 angles) or dense (128 angles) measurements from single or multiple incident waves.

Demonstrate the model's efficacy across various scenarios. For classification, the CNN achieved up to 99.2% accuracy on test sets, with star-shaped obstacles classified perfectly and robust performance under noise levels up to 5%. In regression tasks, peanut-shaped obstacles saw R² scores of 99.92% and RMSE of 0.0594, while kite-shaped ones scored 99.70% R² with RMSE of 0.0970, both maintaining stability in noisy conditions. Star-shaped obstacles posed greater s; with fixed impedance, the model achieved 99.14% R² and RMSE of 0.0159 using deeper networks and more data, whereas variable impedance cases required additional incident waves and attention mechanisms, yielding 96.02% R² and RMSE of 0.0617. Misclassified obstacles near decision boundaries were still reconstructed accurately, highlighting 's resilience.

Of this research are profound for applied sciences and engineering. By enabling precise reconstruction of object shapes and properties from limited far-field data, it advances non-destructive testing, medical diagnostics, and antenna design. The use of circular CNNs tailored to periodic data sets a new standard for handling inverse problems in wave-based imaging, potentially reducing reliance on invasive techniques. Moreover, the divide-and-conquer framework showcases how AI can decompose complex tasks into manageable steps, improving efficiency and accuracy in scenarios where traditional s like boundary integral equations struggle with initial guesses and geometric complexities.

Despite its successes, the approach has limitations. The inverse problem's ill-posedness means uniqueness isn't guaranteed, and performance degrades with higher noise, especially for star-shaped obstacles with variable impedance. Recovery of low-order boundary coefficients like α₁ and β₁ remains challenging due to their subtle influence on scattered fields. Future work could explore non-constant impedance functions, unknown material parameters, or phase-less data, requiring even more expressive neural models. Theoretical analysis of convergence and stability for these circular CNNs is also needed to solidify their foundation in operator learning for inverse problems.

Reference: Mindrinos, L., Pallikarakis, N., and Tsitsas, N.L. (2025). Inverse Electromagnetic Scattering for Doubly-Connected Cylinders using Convolutional Neural Networks. arXiv:2511.20681v1 [math.NA].