BIMSA >
Data Analysis and Problem Solving Seminar
Data Analysis and Problem Solving Seminar
Uncertainty Quantification in Physics-Informed Neural Networks: Quantifying Uncertainty in Predictive Results
Uncertainty Quantification in Physics-Informed Neural Networks: Quantifying Uncertainty in Predictive Results
Organizer
Speaker
Qian Zhang
Time
Friday, July 3, 2026 3:00 PM - 4:00 PM
Venue
A3-1-301
Online
Zoom 204 323 0165
(BIMSA)
Abstract
Physics-informed neural networks provide a flexible framework for solving forward and inverse problems governed by differential equations, but their predictions can be unreliable when training data are noisy, sparse, or when the assumed physical model is incomplete. This talk introduces methods for quantifying uncertainty in PINN predictions from two perspectives: data uncertainty and model uncertainty. We first discuss why a single deterministic prediction is insufficient, and then review several representative approaches, including Bayesian PINNs, Monte Carlo Dropout, deep ensembles, noisy input-output modeling, and discrepancy-based model correction. The talk emphasizes how these methods estimate prediction intervals, distinguish different sources of predictive uncertainty, and improve the reliability of PINNs in scientific modeling.
Speaker Intro
Zhang Qian is a second-year Ph.D. student in a joint Program between BIMSA and Renmin University of China, majoring in Statistics and Big Data, under the supervision of Professor Zhang Xiaoming. Her current research focuses on the data-driven discovery of differential equations from noisy and irregularly sampled time-series data using machine learning techniques.