报告题目 Title: Neural ODEs: Approximation Theory and Operator Learning

报告人 Speaker: 李兹谦

报告人所在单位 Affiliation: 吉林大学、埃尔朗根-纽伦堡大学

时间 Time: 2026-03-02 16:15-17:00

地点 Venue: 光华楼东主楼1801

报告摘要 Abstract: Neural Ordinary Differential Equations (NODEs) have emerged as a powerful paradigm for modeling continuous-time systems. This talk explores the theoretical foundations and practical applications of NODEs in scientific machine learning through two recent works. First, we introduce semi-autonomous NODEs (SA-NODEs) for approximating dynamical systems, establishing their universal approximation properties and quantitative convergence rates with significantly reduced model complexity. Second, we present NODE-ONet, a novel operator learning framework for solving PDEs. By integrating physics-encoded NODEs into an encoder-decoder architecture, NODE-ONet effectively decouples spatial and temporal variables, enabling highly efficient numerical predictions and robust extrapolation beyond training time frames.

个人简介 Bio: 李兹谦，吉林大学博士研究生（导师：汤涛教授、张然教授）、德国埃尔朗根-纽伦堡大学博士研究生（导师：Enrique Zuazua教授）。研究方向包括深度学习求解微分方程、流体混合的最优控制理论与数值计算。相关成果发表或接收在SINUM, M3AS, JSC等期刊上。

海报 Poster: 李兹谦 学术报告.jpg