Physical-Optics-Inspired Machine Learning Scheme for Analysis of Electrically Large Radome

Analysis of radome effects on the antenna scanning performance constitutes a critical aspect in radar line-of-sight calibration and boresight error compensation. For an electrically large antenna-radome system, the traditional full-wave methods are accurate but inefficient, whereas high-frequency methods are efficient but less accurate. In this paper, we present a novel machine learning scheme to bridge the gap between the two approaches. Through training the neural network (NN) that learns the residual of the equivalent currents on the radome surfaces calculated by the method of moments (MoM) and physical optics (PO) method, the proposed scheme can effectively correct the error of PO in evaluating the radiated far field with minimal overhead. It needs only a small amount of accurate data calculated by MoM at a few fixed scanning angles for the correction. Despite counting the time spent on dataset construction and NN training, the approach can save much time compared to MoM. The data processing and NN architecture are designed to preserve the natural symmetries inherent in PO, which equip the scheme with a strong generalization capability. Numerical experiments demonstrate that the trained NN is still reliable when the radome is modified in shape and size.