4 March 2019 in IEEE Transactions on Antennas and Propagation
In this work, we study the application of the supervised descent method (SDM) for two-dimensional microwave imaging. SDM contains offline training and online prediction. In the offline stage, a training data set is generated according to prior information. Then the average descent directions between a fixed initial model and the training models can be learned by iterative schemes. In the online stage, model reconstruction is achieved through iterations based on learned descent directions. This scheme offers a new perspective to incorporate prior information into inversion, and reduce the computational complexity in the online inversion. Synthetic examples validate the accuracy and efficiency of this method.
9 August 2021 in IEEE Transactions on Antennas and Propagation
In this work, we design an iterative deep neural network to solve full-wave inverse scattering problems (ISPs) in the 2-D case. Forward modeling neural networks that predict the scattered field are embedded in an inversion neural network. In an iterative manner, the inversion network predicts the model update from the residual between the simulated data and the observed data. The proposed inversion network can achieve super-resolution reconstruction meanwhile keeping the simulated data of reconstructed models well consistent with the observed data. We validate this method with both synthetic and experimental data inversion. Results show that the inversion network can predict models with high accuracy, efficiency, and good generalization ability. By combining deep learning and physical simulation together, the proposed method provides a way for real-time imaging with high reliability and accuracy.
6 April 2021 in IEEE Transactions on Antennas and Propagation
The volume integral equation (VIE) that describes the forward scattering problem is generally solved by iterative methods, such as the conjugate gradient (CG) method. In this work, we unfold the CG method into an iterative deep neural network to accelerate solving the VIE. After the dielectric scatterer’s relative permittivity and the incident field are input into the network, the total field is trained to converge to the ground truth iteratively. In the neural network, Green’s function is taken as an explicit operator to describe wave physics, and the fast Fourier transform (FFT) is applied to accelerate the computation of volume integrations. The global influence of all points in the space is compressed into a layer by the volume integration. In numerical tests, we validate the accuracy, efficiency, and generalization ability of the proposed neural network, and investigate the feasibility of changing the input size and the frequency in the prediction. Results show that the network is scale-independent and adaptable to predict fields in a narrow frequency band. This work provides us a new perspective of incorporating both learned parameters and physics into numerical algorithms for fast computation, and has the potential of being applied in deep-learning-based inverse scattering problems.
1 March 2023 in IEEE Transactions on Antennas and Propagation
In this study, physics-informed supervised residual learning (PhiSRL) is proposed to enable an effective, robust, and general deep learning framework for 2D electromagnetic (EM) modeling. Based on the mathematical connection between the fixed-point iteration method and the residual neural network (ResNet), PhiSRL aims to solve a system of linear matrix equations. It applies convolutional neural networks (CNNs) to learn updates of the solution with respect to the residuals. Inspired by the stationary and non-stationary iterative scheme of the fixed-point iteration method, stationary and non-stationary iterative physics-informed ResNets (SiPhiResNet and NiPhiResNet) are designed to solve the volume integral equation (VIE) of EM scattering. The effectiveness and universality of PhiSRL are validated by solving VIE of lossless and lossy scatterers with the mean squared errors (MSEs) converging to ~ 10 -4 (SiPhiResNet) and ~ 10 -7 (NiPhiResNet). Numerical results further verify the generalization ability of PhiSRL.
7 January 2019 in IEEE Journal on Multiscale and Multiphysics Computational Techniques
In this work, we study a joint inversion algorithm to reconstruct acoustic and electromagnetic data in biomedical imaging. This algorithm is based on contrast source inversion algorithm (CSI). We define contrast source and contrast functions in both acoustics and electromagnetics, respectively, and apply two reciprocal regularization operators to link acoustic and electromagnetic inversion. By minimizing the cost function using the conjugate gradient method, we can simultaneously reconstruct compressibility, attenuation, permittivity and conductivity of different human tissues. Numerical experiments show that acoustic and electromagnetic inversion can compensate each other and achieve a better reconstruction than individual inversion. Besides, the joint inversion method is particularly sensitive to compressibility and conductivity and thus may be used to monitor air and water contents in human thorax as well as other applications in biomedicine.
9 June 2020 in IEEE Transactions on Antennas and Propagation
We present a general framework of applying supervised descent method (SDM) to the pixel-based and model-based full-wave inversion of microwave data for dielectric targets. SDM is a machine learning approach that utilizes the learned descent directions to reconstruct models. In this paper, we study the pixel-based inversion with an online regularization scheme, and the model-based inversion with the parametric level set approach. Furthermore, an online restart scheme is studied to further reduce the data residual. In addition, we investigate the generalization ability of this machine-learning algorithm. In the numerical test, the pixel-based SDM inversion are used to process both single-or multi-frequency data, and the model-based inversion are performed on data with limited observations. Both synthetic and experimental results show good accuracy, speed, and generalization ability of this algorithm. SDM may provide us a potential way to improve reconstruction quality in nonlinear inversion through information fusion of measured data, microwave physics, and various prior information.