Improved physics-informed neural network in mitigating gradient-related failures

被引：2

作者：

Niu, Pancheng ^{[1
]}

Guo, Jun ^{[1
,2
]}

Chen, Yongming ^{[1
]}

Zhou, Yuqian ^{[1
]}

Feng, Minfu ^{[3
]}

Shi, Yanchao ^{[4
]}

机构：

[1] Chengdu Univ Informat Technol, Coll Appl Math, Chengdu 610225, Sichuan, Peoples R China

[2] Neijiang Normal Univ, Key Lab Numer Simulat Sichuan Prov Univ, Sch Math & Informat Sci, Neijiang 641000, Sichuan, Peoples R China

[3] Sichuan Univ, Coll Math, Chengdu 610064, Sichuan, Peoples R China

[4] Southwest Petr Univ, Sch Sci, Chengdu 610500, Sichuan, Peoples R China

来源：

NEUROCOMPUTING | 2025年 / 638卷

基金：

中国国家自然科学基金;

关键词：

Physics-informed neural networks; Gradient flow stiffness; Adaptive weighting; Scientific computing; DEEP LEARNING FRAMEWORK;

D O I：

10.1016/j.neucom.2025.130167

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

Physics-informed neural network (PINN) integrates fundamental physical principles with advanced data driven techniques, leading to significant advancements in scientific computing. However, PINN encounters persistent challenges related to stiffness in gradient flow, which limits their predictive capabilities. This paper introduces an improved PINN (I-PINN) designed to mitigate gradient-related failures. The core of I-PINN combines the respective strengths of neural networks with an improved architecture and adaptive weights that include upper bounds. I-PINN achieves improved accuracy by at least one order of magnitude and accelerate convergence without introducing additional computational complexity compared to the baseline model. Numerical experiments across a variety of benchmarks demonstrate the enhanced accuracy and generalization of I-PINN. The supporting data and code are accessible at https://github.com/PanChengN/IPINN.git, facilitating broader research engagement.

引用

页数：14

共 31 条

[11] Locally adaptive activation functions with slope recovery for deep and physics-informed neural networks [J].

Jagtap, Ameya D. ;

Kawaguchi, Kenji ;

Karniadakis, George Em .

PROCEEDINGS OF THE ROYAL SOCIETY A-MATHEMATICAL PHYSICAL AND ENGINEERING SCIENCES, 2020, 476 (2239)

[12] Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems [J].

Jagtap, Ameya D. ;

Kharazmi, Ehsan ;

Karniadakis, George Em .

COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING, 2020, 365

[13] Adaptive activation functions accelerate convergence in deep and physics-informed neural networks [J].

Jagtap, Ameya D. ;

Kawaguchi, Kenji ;

Karniadakis, George Em .

JOURNAL OF COMPUTATIONAL PHYSICS, 2020, 404 (404)

[14] Physics-informed machine learning [J].

Karniadakis, George Em ;

Kevrekidis, Ioannis G. ;

Lu, Lu ;

Perdikaris, Paris ;

Wang, Sifan ;

Yang, Liu .

NATURE REVIEWS PHYSICS, 2021, 3 (06) :422-440

[15] Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics [J].

Kendall, Alex ;

Gal, Yarin ;

Cipolla, Roberto .

2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), 2018, :7482-7491

[16]

Krishnapriyan AS, 2021, ADV NEUR IN, V34

[17] Equity-premium prediction: Attention is all you need [J].

Lima, Luiz Renato ;

Godeiro, Lucas Lucio .

JOURNAL OF APPLIED ECONOMETRICS, 2023, 38 (01) :105-122

[18] A Dual-Dimer method for training physics-constrained neural networks with minimax architecture [J].

Liu, Dehao ;

Wang, Yan .

NEURAL NETWORKS, 2021, 136 :112-125

[19] PHYSICS-INFORMED NEURAL NETWORKS WITH HARD CONSTRAINTS FOR INVERSE DESIGN\ast [J].

Lu, Lu ;

Pestourie, Raphael ;

Yao, Wenjie ;

Wang, Zhicheng ;

Verdugo, Francesc ;

Johnson, Steven G. .

SIAM JOURNAL ON SCIENTIFIC COMPUTING, 2021, 43 (06) :B1105-B1132

[20] DeepXDE: A Deep Learning Library for Solving Differential Equations [J].

Lu, Lu ;

Meng, Xuhui ;

Mao, Zhiping ;

Karniadakis, George Em .

SIAM REVIEW, 2021, 63 (01) :208-228

← 1 2 3 4 →