Uncertainty Quantification and Deep Learning for Physical Problems

Speaker:    Prof. Guang Lin

Time:        15:00-16:00, June 13

Location:    SIST 1A-108

Host:       Prof. Qifeng Liao

Abstract:

In this talk, we will present three new data-driven, deep-learning based methods for predicting outcomes of complex nonlinear physical systems and quantifying the uncertainties in deep learning. First, we will introduce a fast probabilistic convolutional encoder-decoder network named PDE-UQ-Net for predicting the solutions of heterogeneous elliptic partial differential equations on varied domains. Unlike other approaches, PDE-UQ-Net can quantify the uncertainties in deep-learning-based prediction and allow training to be scaled to the large data sets. In addition, to predict material failure and fracture propagation, we will present a new deep neural network named Peri-Net. Finally, to efficiently predict the solutions for heterogeneous nonlinear multiphase flow problems, we will introduce a multiscale deep neural network named MS-Net. We will demonstrate the power and efficiency of the above three neural networks in solving complex physical problems by comparing with traditional methods.

Bio:

林光教授现任美国普渡大学机械工程学院和数学系的博师生导师和终身副教授，普渡大学数据科学咨询中心主任。2007年，他在美国常春藤名校-布朗大学获得了应用数学博士学位。在他被任命为普渡大学教授之前，他曾在2007年至2014年在美国能源部太平洋西北国家实验室担任高性能计算、数学和数据学习部门的资深研究员。

同时林光教授目前还在国际不确定度量化杂志、奥斯汀统计杂志、随机学杂志，和科学世界杂志的编辑委员会任编委。

SIST-Seminar 18173