構造工学おける物理インフォームドニューラルネットワーク
氏名：金 哲佑
所属：京都大学
概要：In structural engineering, recent advances have increasingly integrated physics and data to enable fast and accurate modeling of complex mechanical systems. The studies highlight two representative directions: physics-informed machine learning for mesh-free static solid mechanics and data-driven neural operator for vehicle-bridge interaction system (VBI). These demonstrate the potential of solvers for efficient structural analysis and monitoring.
1. The Multi-resolution Hash-encoding Solver (MHS) has been developed to enhance computational efficiency and accuracy in solving solid mechanics. As a pure physics-driven alternative, MHS utilizes the sampled point cloud as input and outputs the query point cloud state with exactly satisfied Dirichlet boundary conditions. Compared to vanilla PINN and EPINN, MHS obtained the solution fields much faster and more accurately, achieving a 40-fold speedup in the hyperelastic cube case.
2. The two-dimensional vehicle–bridge Interaction Neural Operator (VINO2D) framework extends VINO to spatio-temporal PDE learning of a complex coupled vehicle–bridge interaction system. The model predicts the distribution of bridge damage from structural response fields, enabling rapid damage detection, localization, and quantification within a digital twin framework. The results show accurate inverse identification with high computational efficiency, demonstrating strong potential for bridge structural health monitoring.
所属：京都大学
概要：In structural engineering, recent advances have increasingly integrated physics and data to enable fast and accurate modeling of complex mechanical systems. The studies highlight two representative directions: physics-informed machine learning for mesh-free static solid mechanics and data-driven neural operator for vehicle-bridge interaction system (VBI). These demonstrate the potential of solvers for efficient structural analysis and monitoring.
1. The Multi-resolution Hash-encoding Solver (MHS) has been developed to enhance computational efficiency and accuracy in solving solid mechanics. As a pure physics-driven alternative, MHS utilizes the sampled point cloud as input and outputs the query point cloud state with exactly satisfied Dirichlet boundary conditions. Compared to vanilla PINN and EPINN, MHS obtained the solution fields much faster and more accurately, achieving a 40-fold speedup in the hyperelastic cube case.
2. The two-dimensional vehicle–bridge Interaction Neural Operator (VINO2D) framework extends VINO to spatio-temporal PDE learning of a complex coupled vehicle–bridge interaction system. The model predicts the distribution of bridge damage from structural response fields, enabling rapid damage detection, localization, and quantification within a digital twin framework. The results show accurate inverse identification with high computational efficiency, demonstrating strong potential for bridge structural health monitoring.
論文掲載,発表実績:
（学術雑誌掲載論文）
- C. Kaewnuratchadasorn, J. Wang, C.W. Kim, Y. Yang, "Two-Dimensional Vehicle–Bridge Interaction Neural Operator for Digital Twin of Bridge Structures", Structural Control and Health Monitoring, 2025(1): 1031214, 2025.
- X. Tian, J. Wang, X. Lu, "Exact Dirichlet boundary multi‐resolution hash encoding solver for structures", Computer-Aided Civil and Infrastructure Engineering, 40(25), 4172–4192, 2025.
Posted : 2026年03月31日
