EI、Scopus 收录
中文核心期刊
李道伦, 沈路航, 查文舒, 邢燕, 吕帅君, 汪欢, 李祥, 郝玉祥, 陈东升, 陈恩源. 基于神经算子与类物理信息神经网络智能求解新进展. 力学学报, 2024, 56(3): 1-15. DOI: 10.6052/0459-1879-23-407
引用本文: 李道伦, 沈路航, 查文舒, 邢燕, 吕帅君, 汪欢, 李祥, 郝玉祥, 陈东升, 陈恩源. 基于神经算子与类物理信息神经网络智能求解新进展. 力学学报, 2024, 56(3): 1-15. DOI: 10.6052/0459-1879-23-407
Li Daolun, Shen Luhang, Zha Wenshu, Xing Yan, Lyu Shuaijun, Wang Huan, Li Xiang, Hao Yuxiang, Chen Dongsheng, Chen Enyuan. New progress in intelligent solution of neural operators and physics-informed-based methods. Chinese Journal of Theoretical and Applied Mechanics, 2024, 56(3): 1-15. DOI: 10.6052/0459-1879-23-407
Citation: Li Daolun, Shen Luhang, Zha Wenshu, Xing Yan, Lyu Shuaijun, Wang Huan, Li Xiang, Hao Yuxiang, Chen Dongsheng, Chen Enyuan. New progress in intelligent solution of neural operators and physics-informed-based methods. Chinese Journal of Theoretical and Applied Mechanics, 2024, 56(3): 1-15. DOI: 10.6052/0459-1879-23-407

基于神经算子与类物理信息神经网络智能求解新进展

NEW PROGRESS IN INTELLIGENT SOLUTION OF NEURAL OPERATORS AND PHYSICS-INFORMED-BASED METHODS

  • 摘要: 深度学习通过多层神经网络对数据进行学习, 不仅能揭示潜藏信息, 还能很好地解决复杂非线性问题. 偏微分方程(PDE)是描述自然界中许多物理现象的基本数学模型. 两者的碰撞与融合, 产生了基于深度学习的PDE智能求解方法, 它具有高效、灵活和通用等优点. 文章聚焦PDE智能求解方法, 以是否求解单一问题为判定依据, 把求解方法分为两类: 神经算子方法和类物理信息神经网络(PINN)方法, 其中神经算子方法用于求解一类具有相同数学特征的PDE问题, 类PINN方法用于求解单一问题. 对于神经算子方法, 从数据驱动和物理约束两个方面展开介绍, 分析研究现状并指出现有方法的不足. 对于类PINN方法, 首先介绍了基础PINN的3种改进方法(基于数据优化、基于模型优化和基于领域知识优化), 然后详细介绍了基于物理驱动的两类解决方案: 基于传统PDE离散方程的智能求解方案和无网格的非离散求解方案. 最后总结技术路线, 探讨现有研究存在的不足, 给出可行的研究方案. 最后, 简要介绍智能求解程序发展现状, 并对未来研究方向给出建议.

     

    Abstract: Deep learning, characterized by its multi-layer neural networks, has demonstrated its capability not only to uncover hidden information in data but also to effectively address complex nonlinear problems. As a fundamental mathematical model, partial differential equations (PDEs) find wide application in describing various physical phenomena in the natural world. The amalgamation of deep learning and PDEs has given rise to the emergence of intelligent PDE solving methods based on deep learning. These methods possess several advantageous traits, including high efficiency, flexibility, and universality, which make them valuable in practical applications. This paper focuses on intelligent PDE solving methods, categorizing the solving approaches into two types based on whether they handle single or multiple problems: neural operator methods and physics-informed neural network (PINN) methods. Neural operator methods are employed to solve a class of PDE problems with the same mathematical characteristics, while PINN-based methods are used to solve single problems. The first category encompasses neural operator methods, which are utilized for solving a group of PDE problems that share similar mathematical characteristics. These methods leverage data-driven approaches and physical-constraint approaches to formulate their solutions. An in-depth analysis is conducted to examine the current research status, along with the identification of existing drawbacks in these approaches. As for PINN-based methods, this paper introduces the relevant research progress from three derivative approaches of PINN (data-optimization, model-optimization, and domain-knowledge-optimization). Finally, this paper provides a comprehensive overview of the technical roadmap for PDE intelligent solving methods. It critically evaluates the existing research limitations and proposes feasible research plans to overcome these challenges. Additionally, the paper briefly introduces the current state of intelligent solving program development and offers suggestions for future research directions. By amalgamating deep learning and PDEs, these intelligent solving methods have the potential to revolutionize various scientific and engineering domains, enabling more accurate and efficient problem-solving in complex nonlinear systems.

     

/

返回文章
返回