Uncertainty Nonlinearly Guided Learning Framework for Full-Wave Inverse Scattering

Recently, deep learning schemes have been widely used in electromagnetic society. However, different from traditional methods constrained by physics models, the predicted results of these “black-box” neural networks are not reliable, and the prediction errors of networks can only be obtained when ground truths are known, which severely limits the applications of learning-based solvers. Consequently, uncertainty quantifications (UQs) are introduced in learning-based electromagnetic solvers to provide a confidence level. In this work, by incorporating the properties of physical setup in inverse scattering problems (ISPs), an equivariance-based uncertainty quantification (EUQ) method is first proposed, where the diversities caused by different antennas are utilized to quantify the model uncertainty with virtual experiments. More importantly, the quantified uncertainties further serve as nonlinear guidances in ISP learning solvers, which makes it flexible to adaptively choose targeted reconstruction regions. It is shown by intensive tests that the proposed EUQ has a better correlation with the true error compared with traditional UQ methods. The learning framework nonlinearly guided by EUQ also shows much better performance and achieves better predictions in target selected areas in both numerical and experimental tests.