报告人：周泽慧博士

报告题目：Structured Neural Networks for Inverse Scattering: An Explainable Approach

报告摘要：Recently, deep neural networks (DNNs) have emerged as powerful tools for solving inverse scattering problems. However, their interpretability, as well as their approximation and generalization properties in this context, remain largely under-explored.

In this talk, I will present an explainable neural network framework for the simultaneous recovery of two function-valued coefficients in the Helmholtz equation using multi-frequency scattering data. The network design is motivated by a decomposition of the regularized pseudo-inverse of the linearized forward operator into two components: a Fourier integral operator and a convolution-type inverse. Each component is approximated by a dedicated neural module that mirrors its mathematical structure—shared-weight fully connected layers for the Fourier component, and residual convolutional layers for the convolution-type inverse. I will also present theoretical results on the approximation and generalization capabilities of the proposed networks, supported by numerical experiments.

报告时间：2025年7月16日 14:00 – 16:00

报告地点：正心楼111

报告人简介：周泽慧，香港中文大学博士后研究人员。2018年毕业于武汉大学数学与统计学院获学士学位，2022年毕业于香港中文大学数学系获博士学位，2022-2025年于美国新泽西州里罗格斯大学担任Hill助理教授。研究方向为大规模反问题的高效数值算法和可解释的深度学习方法，相关成果发表在SIAM J. Optim.，SIAM J. Imaging Sci.，SIAM/ASA J. Uncertain. Quantif.，Inverse Probl.，Neural Netw. 等期刊上。