报告题目：Explicit-Interface Physics-Informed Neural Networks for Hyperbolic Conservation Laws

报告人：赵张驰

报告时间：2026年9月28日下午，4:30-5:30

地点：管理科研楼1218室

摘要： 

Hyperbolic conservation laws often develop shocks and other discontinuous structures, posing fundamental difficulties for standard physics-informed neural networks based on continuous function representations and pointwise strong-form residuals. This report presents an explicit-interface physics-informed neural framework for conservation laws with shocks. The solution is represented by two neural branches describing the smooth states on either side of the shock, together with a learnable interface that explicitly tracks the shock position and speed. Governing equations are enforced separately in the smooth regions, while the Rankine–Hugoniot condition couples the one-sided states across the interface. Entropy and Lax conditions are further incorporated to ensure physical admissibility and select the appropriate shock family. Numerical experiments on scalar conservation laws, one-dimensional Euler shock-tube problems, and two-dimensional Euler flows demonstrate the ability of the proposed framework to preserve sharp discontinuities while accurately tracking shock motion.

报告人简介：

赵张弛，西安交通大学统计学博士研究生，导师为徐宗本院士，主要从事深度学习与元学习算法研究。相关研究成果发表于ICML 2025。