Enhancing training of physics-informed neural networks using domain-decomposition based preconditioning strategies

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Kopaničáková, Alena, Kothari, Hardik, Karniadakis, George Em, Krause, Rolf
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909168432054272
author Kopaničáková, Alena
Kothari, Hardik
Karniadakis, George Em
Krause, Rolf
author_facet Kopaničáková, Alena
Kothari, Hardik
Karniadakis, George Em
Krause, Rolf
contents We propose to enhance the training of physics-informed neural networks (PINNs). To this aim, we introduce nonlinear additive and multiplicative preconditioning strategies for the widely used L-BFGS optimizer. The nonlinear preconditioners are constructed by utilizing the Schwarz domain-decomposition framework, where the parameters of the network are decomposed in a layer-wise manner. Through a series of numerical experiments, we demonstrate that both, additive and multiplicative preconditioners significantly improve the convergence of the standard L-BFGS optimizer, while providing more accurate solutions of the underlying partial differential equations. Moreover, the additive preconditioner is inherently parallel, thus giving rise to a novel approach to model parallelism.
format Preprint
id arxiv_https___arxiv_org_abs_2306_17648
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Enhancing training of physics-informed neural networks using domain-decomposition based preconditioning strategies
Kopaničáková, Alena
Kothari, Hardik
Karniadakis, George Em
Krause, Rolf
Numerical Analysis
Machine Learning
Optimization and Control
90C30, 90C26, 90C06, 65M55, 68T07
We propose to enhance the training of physics-informed neural networks (PINNs). To this aim, we introduce nonlinear additive and multiplicative preconditioning strategies for the widely used L-BFGS optimizer. The nonlinear preconditioners are constructed by utilizing the Schwarz domain-decomposition framework, where the parameters of the network are decomposed in a layer-wise manner. Through a series of numerical experiments, we demonstrate that both, additive and multiplicative preconditioners significantly improve the convergence of the standard L-BFGS optimizer, while providing more accurate solutions of the underlying partial differential equations. Moreover, the additive preconditioner is inherently parallel, thus giving rise to a novel approach to model parallelism.
title Enhancing training of physics-informed neural networks using domain-decomposition based preconditioning strategies
topic Numerical Analysis
Machine Learning
Optimization and Control
90C30, 90C26, 90C06, 65M55, 68T07
url https://arxiv.org/abs/2306.17648