11112223333

科学计算系列学术报告:Deep adaptive sampling for numerical PDEs

发布人:日期:2022年11月15日 14:48浏览数:

报告题目:Deep adaptive sampling for numerical PDEs

报 告 人:周涛研究员(中国科学院数学与系统科学研究院)

报告时间:20221117日  15:30

报告地点:腾讯会议(747391985

报告摘要:

Adaptive computation is of great importance in numerical simulations. The ideas for adaptive computations can be dated back to adaptive finite element methods in 1970s. In this talk, we shall first review some recent development for adaptive method with applications. Then, we shall propose a deep adaptive sampling method for solving PDEs where deep neural networks are utilized to approximate the solutions. In particular, we propose the failure informed PINNs (FI-PINNs), which can adaptively refine the training set with the goal of reducing the failure probability. Compared to the neural network approximation obtained with uniformly distributed collocation points, the developed algorithms can significantly improve the accuracy, especially for low regularity and high-dimensional problems.

报告人简介:

周涛,中国科学院数学与系统科学研究院研究员。曾于瑞士洛桑联邦理工大学从事博士后研究。主要研究方向为不确定性量化、偏微分方程数值方法以及时间并行算法等,近期的主要研究课题包括深度学习与科学计算,深度生成模型及其应用等。在国际权威期刊如SIAM ReviewSINUMJCP等发表论文70篇。2018年获优秀青年科学基金资助,2022年获中组部高层次人才专项资助。现担任SIAM J. Sci. Comput.J. Sci. Comput.等国际期刊编委,国际不确定性量化期刊(International Journal for UQ副主编。

上一条:分析系列学术报告:On the connected components of IFS fractals

下一条:科学计算系列学术报告:Exponential convergence of hp-discontinuous Galerkin method for nonlinear Caputo fractional differential equations

【关闭】 打印    收藏