ConvPDE-UQ: Convolutional neural networks with quantified uncertainty for heterogeneous elliptic partial differential equations on varied domains

被引:63
作者
Winovich, Nick [1 ]
Ramani, Karthik [2 ]
Lin, Guang [1 ,2 ]
机构
[1] Purdue Univ, Dept Math, 150 N Univ St, W Lafayette, IN 47907 USA
[2] Purdue Univ, Sch Mech Engn, 585 Purdue Mall, W Lafayette, IN 47907 USA
基金
美国国家科学基金会;
关键词
Partial differential equations; Uncertainty quantification; Convolutional encoder-decoder networks; Deep learning; Machine learning; Confidence interval; OPTIMIZATION; REPRESENTATION; FRAMEWORK; ALGORITHM;
D O I
10.1016/j.jcp.2019.05.026
中图分类号
TP39 [计算机的应用];
学科分类号
081203 ; 0835 ;
摘要
In this work, we introduce the ConvPDE-UQ framework for constructing light-weight numerical solvers for partial differential equations (PDEs) using convolutional neural networks. A theoretical justification for the neural network approximation to partial differential equation solvers on varied domains is established based on the existence and properties of Green's functions. These solvers are able to effectively reduce the computational demands of traditional numerical methods into a single forward-pass of a convolutional network. The network architecture is also designed to predict pointwise Gaussian posterior distributions, with weights trained to minimize the associated negative log-likelihood of the observed solutions. This setup facilitates simultaneous training and uncertainty quantification for the network's solutions, allowing the solver to provide pointwise uncertainties for its predictions. The associated training procedure avoids the computationally expensive Bayesian inference steps used by other state-of-the-art uncertainty models and allows training to be scaled to the large data sets required for learning on varied problem domains. The performance of the framework is demonstrated on three distinct classes of PDEs consisting of two linear elliptic problem setups and a nonlinear Poisson problem. After a single offline training procedure for each class, the proposed networks are capable of accurately predicting the solutions to linear and nonlinear elliptic problems with heterogeneous source terms defined on any specified two-dimensional domain using just a single forward-pass of a convolutional neural network. Additionally, an analysis of the predicted pointwise uncertainties is presented with experimental evidence establishing the validity of the network's uncertainty quantification schema. (C) 2019 Elsevier Inc. All rights reserved.
引用
收藏
页码:263 / 279
页数:17
相关论文
共 59 条
  • [1] Unified framework for finite element assembly
    Alnaes, M. S.
    Logg, A.
    Mardal, K-A.
    Skavhaug, O.
    Langtangen, H. P.
    [J]. INTERNATIONAL JOURNAL OF COMPUTATIONAL SCIENCE AND ENGINEERING, 2009, 4 (04) : 231 - 244
  • [2] Unified Form Language: A Domain-Specific Language for Weak Formulations of Partial Differential Equations
    Alnaes, Martin S.
    Logg, Anders
    Olgaard, Kristian B.
    Rognes, Marie E.
    Wells, Garth N.
    [J]. ACM TRANSACTIONS ON MATHEMATICAL SOFTWARE, 2014, 40 (02):
  • [3] On the Efficiency of Symbolic Computations Combined with Code Generation for Finite Element Methods
    Alnaes, Martin Sandve
    Mardal, Kent-Andre
    [J]. ACM TRANSACTIONS ON MATHEMATICAL SOFTWARE, 2010, 37 (01):
  • [4] Alnaes MS., 2015, Archive of numerical software, V3, P1, DOI [10.11588/ans. 2015.100.20553, DOI 10.11588/ANS.2015.100.20553]
  • [5] [Anonymous], UFC FINITE ELEMENT C
  • [6] [Anonymous], 2012, Lecture Notes in Computational Science and Engineering
  • [7] [Anonymous], FFC FENICS FORM COMP
  • [8] [Anonymous], ARXIV13126114
  • [9] [Anonymous], 2017, P ADV NEUR INF PROC
  • [10] [Anonymous], UFL FINITE ELEMENT F