Towards Calibration of Uncooperative Electromagnetic Imaging Systems using CycleGANs

An uncooperative electromagnetic inversion system, e.g., a grain storage bin, makes it difficult to calibrate the numerical and physical system models. In this work, we investigate the idea of training a CycleGAN to convert experimentally collected S-parameter data into calibrated field data suitable for subsequent electromagnetic inversion. A simple proof of concept, applied to synthetically generated "experimental" data, is implemented, trained, and evaluated. Results demonstrate the potential for CycleGANs to play an important role in electromagnetic imaging.