EI、Scopus 收录
中文核心期刊

物理驱动深度学习波动数值模拟方法及应用

METHODS AND APPLICATIONS OF PHYSICAL INFORMATION DEEP LEARNING IN WAVE NUMERICAL SIMULATION

  • 摘要: 近年来, 物理先验融合数据的深度学习方法求解以偏微分方程为理论基础的正反演问题已成为交叉学科热点. 针对地震工程波动数值模拟, 本文阐明了物理驱动深度学习方法PINN的数学概念及实现方式, 以无源项一维波动为例, 开展了相关理论模型构建, 并与解析解及有限差分方法进行对比, 分析了PINN方法与其他数值算法模拟波场的相对范数误差, 验证了物理驱动深度学习方法求解波动问题的可行性. 采用物理驱动深度学习方法并结合谱元法形成的稀疏初始波场数据, 开展了二维波动数值模拟, 实现了自由边界条件及起伏地表等典型工况的模拟, 并给出了时序波场分布特性. 更换不同的初始条件, 测试了神经网络的泛化精度, 提出可显著提高网络训练效率的迁移学习方法. 通过与谱元法的结果对比, 验证了本文方法模拟均质场地、空间不均匀及复杂地形场地波动问题的可靠性. 结果表明, 物理驱动深度学习方法具备无网格、精细化模拟等优势, 并可实现自由地表及侧边界波场透射等数值模拟条件.

     

    Abstract: In recent years, physics-informed deep learning methods based on prior data fusion to solve forward and inverse problems based on partial differential equations (PDEs) have become a cross-disciplinary hotspot. This paper clarifies the mathematical concept and implementation of physics-informed neural networks (PINN) for the earthquake engineering numerical simulation of waves. Taking the one-dimensional fluctuation of passive term as an example, the relevant theoretical model of PINN is constructed. The feasibility of physics-driven deep learning methods in solving fluctuation problems is verified by comparing with analytical solutions and finite difference methods. The relative \mathcalL_2 norm errors of the wave field simulated by PINN method and other numerical algorithms are analyzed. The physical driven deep learning method combined with sparse initial wave field data formed by spectral element method is used to numerically simulate two-dimensional fluctuation problem. Typical working conditions such as free boundary conditions and undulating ground surface are realized, and the distribution characteristics of time series wave field are given. Different initial conditions are changed to test the generalization accuracy of the neural network, and a transfer learning method was proposed to significantly improve the training efficiency of the network. By using transfer learning, wave fields at different source locations in infinite media can be directly predicted with high accuracy. Comparing with the results of spectral element method, the reliability of the proposed method is verified to simulate the wave propagation of homogeneous site, spatial inhomogeneity and complex terrain site fluctuation. The results show that the physical-driven deep learning method has the advantages of meshless and fine-grained numerical simulation, and can realize the numerical simulation conditions such as free surface and side/bottom boundary wave field transmission.

     

/

返回文章
返回