BIMSA >
Data Analysis and Problem Solving Seminar
Data Analysis and Problem Solving Seminar
Training dynamics in physics-informed machine learning
Training dynamics in physics-informed machine learning
Organizer
Speaker
Yuzhong Hu
Time
Friday, July 31, 2026 3:00 PM - 4:00 PM
Venue
A3-1-301
Online
Zoom 204 323 0165
(BIMSA)
Abstract
Physics-informed neural networks (PINNs) provide a flexible framework for solving forward and inverse problems governed by differential equations, but their training can be slow, unstable, and strongly imbalanced across different loss components.
This talk reviews recent theoretical insights into PINN training dynamics from two perspectives. First, the neural tangent kernel (NTK) framework is used to explain spectral bias and the unequal convergence rates of PDE, boundary, and data losses. Second, information bottleneck theory and the gradient signal-to-noise ratio (SNR) are introduced to describe three training stages: fitting, diffusion, and total diffusion.
This talk reviews recent theoretical insights into PINN training dynamics from two perspectives. First, the neural tangent kernel (NTK) framework is used to explain spectral bias and the unequal convergence rates of PDE, boundary, and data losses. Second, information bottleneck theory and the gradient signal-to-noise ratio (SNR) are introduced to describe three training stages: fitting, diffusion, and total diffusion.
Speaker Intro
Hu Yuzhong is a second-year PhD student in a joint program between BIMSA and academy of mathematics and systems science, CAS, under the supervision of Professor Zhang Xiaoming.