A Mesh-Free, Broadband, Multi-Input Intelligent RCS Prediction Method Based on PointNet++

This article presents an artificial intelligence (AI) method for efficiently predicting the scattering field of 3-D perfect-electrical-conducting (PEC) objects with random structures under varying frequencies and angles. A point cloud-based, mesh-free method eliminates the need for meshing process, substantially reducing computational time and resources. The proposed framework consists of two components corresponding to two types of input data formats. Spatially non-uniformly sampled point clouds represent the geometric features of the targets, serving as the input for the backbone network-PointNet++. Simulation parameters, including frequency and elevation angle, are fed into the auxiliary networks. Compared to other radar cross section (RCS) prediction methods, our model has two distinct advantages: the point cloud data significantly reduces the memory footprint of the dataset, allowing for larger batch sizes during training; the frequency and elevation angle can be flexibly adjusted through the auxiliary network's input channels. Thus, the method achieves a degree of broadband generalization capability. Our proposed method has been validated through numerical experiments and has significantly enhanced efficiency and accuracy. By combining computer vision and computational electromagnetic research, this study makes full use of the flexibility of point cloud data, opening a new vista for the application of deep learning in solving practical engineering problems.