RCS Optimization of Surface Geometry With Physics Inspired Neural Networks

Abstract

Radar cross section (RCS) optimization is important to object geometry design, for example seeking a low-scattering structure. However, it is difficult to obtain a geometry with particular RCS quickly due to the complex geometry, low-efficient RCS calculation, or lack of effective automatic optimization methods. In this paper, a RCS optimization method is proposed based on physics inspired neural network named electromagnetic fully connected neural network (EM-FCNN). It employs the principles of MoM to transform the slow numerical calculation method into the fast neural network calculation. To reduce the complexity of surface geometry characterization, a low-dimensional surface hyperparametric modulation method (SHMM) is formulated to characterize object surfaces by introducing a modulation factor into rough surfaces. In this regard, the ultra-high-dimensional target surfaces can be characterized by only a few hyperparameters. To accelerate the optimization process, a dimensional reduction optimization algorithm (DROA) is further designed to simplify the multi-dimensional hyperparameters optimization problem to a series of one-dimensional optimization problems. The efficacy of the proposed method is validated with a RCS reduction task of a simplified aircraft model. This is generalized to solve the RCS optimization and it can be used to handle object geometry design for other application areas.