An AI Predictor: From Point Clouds to Scattered Far Fields for 3-D PEC Targets

In this article, an artificial intelligence (AI) approach is proposed for the prediction of scattered far fields by 3-D perfect-electrical-conducting (PEC) targets. The conventional computational electromagnetic (CEM) methods are plagued by at least two issues. One is the mesh generation, and the other is the intensive or time-consuming solution of matrix equations which could be unbearable. Machine-learning (ML) methods provide a new perspective to tackle the problems by skipping the two steps. Point clouds can well describe the geometrical information of a radar target. Directly using the point cloud rather than the created mesh with the point data should be workable, so we will take the point cloud as the input of the neural network (NN). As for the output, we will let the NN extract a kind of inherent feature parameters (IFPs) that are distinctive to each target and independent of incident and scattering directions. The characteristic mode theory (CMT) is employed to acquire the IFPs for the generation of datasets. The scattered far-field can be readily computed using the IFPs when the incident and scattering directions are specified. We call the NN constructed to realize the function Electromagnetic PointNet (EMPN). Numerical examples show that the EMPN can achieve remarkable reductions in CPU time and even memory requirements. The proposed approach is convenient to run from the input of a point cloud file that may be obtained by using a cellphone with a depth camera to the output of scattered far-field or radar cross section (RCS).