A Physics-Informed Hierarchical Neural Network for Microwave Scattering Analysis of 3D PEC Targets

Researchers present U-PINet, a physics-informed neural network that accelerates 3D microwave scattering analysis for radar applications by combining graph-based near-field encoding with hierarchical multi-scale fusion, achieving faster computation than classical solvers while maintaining accuracy on complex geometries.

This research addresses a computational bottleneck in electromagnetic modeling where traditional solvers like MLFMA provide high accuracy but become prohibitively slow when analyzing multiple frequencies or incidence angles—a common requirement in radar system design and optimization. The U-PINet architecture ingeniously merges physics-based constraints with machine learning by training on residuals of integral equations rather than requiring labeled current solutions, reducing data annotation overhead while maintaining physical validity.

The approach builds on decades of electromagnetic theory, particularly near-far field decomposition principles that underpin existing fast algorithms, but applies modern deep learning techniques to capture multi-scale interactions through octree-based hierarchical fusion. This represents a meaningful shift toward hybrid surrogate models that leverage domain knowledge rather than purely black-box approaches, addressing a known weakness where purely data-driven methods fail on geometrically unfamiliar targets.

For the aerospace and defense sectors, this technology has immediate practical value. Radar cross-section (RCS) prediction drives antenna and stealth design processes requiring thousands of simulations across parameter spaces. Achieving substantial runtime savings without sacrificing accuracy on complex geometries directly translates to faster design iteration cycles and reduced computational infrastructure costs. The physics-informed training approach suggests the model may generalize better to out-of-distribution geometries than conventional neural surrogate models.

The work demonstrates growing maturity in physics-informed neural networks, moving beyond academic proof-of-concept toward specialized domain applications. Future developments likely include extension to complex materials, dielectric targets, and integration into commercial electromagnetic design software, potentially reshaping computational workflows in RF/microwave engineering.

• →U-PINet combines physics-informed training with hierarchical neural architecture inspired by existing MLFMA solvers for computational efficiency.
• →The method trains on electromagnetic field equation residuals rather than labeled current data, reducing annotation requirements while maintaining physical fidelity.
• →Benchmarks show substantial runtime improvements over classical solvers in repeated-query scenarios common in radar system design.
• →Octree-based multi-scale fusion enables accurate scattering prediction on geometrically complex 3D targets across multiple frequencies and polarizations.
• →Architecture demonstrates how domain-specific knowledge can improve neural network generalization in computational physics applications.