Physics‐Informed Neural Network Analysis of Laterally Loaded Piles Based on Elastic Subgrade Reaction Approach

ABSTRACT

Conventional numerical methods for analyzing laterally loaded piles, such as the finite element method (FEM), involve matrix assembly and inversion, which leads to increased computational overhead for tasks that require massive computations, such as reliability analysis and optimization design. Recently, physics‐informed neural networks (PINN) have emerged as a mesh‐free technique that relies on randomly sampled spatial collocation points to approximate partial differential equations (PDE) solutions. Although this constitutes a form of spatial discretization, it eliminates the need for mesh connectivity, enabling scalable analysis of high‐dimensional problems. Leveraging these advantages, this study establishes a PINN‐based framework for laterally loaded piles that considers multiple input parameters over wide value ranges, including external loads, pile properties and soil properties. A non‐dimensional approach is employed to normalize the input variables, and the linear elastic superposition is utilized to enhance the model's training efficiency and predictive applicability. The model can be trained purely based on PDEs or in combination with data, providing greater flexibility and physical interpretability than purely data‐driven models. Its performance is validated against field measurements and extensive FEM simulations. Applications involving extensive calculations and differentiations are conducted, demonstrating the proposed PINN model's excellent computational efficiency and its advantages in reliability analysis, optimization design, and pile group analysis.