Physics informed neural network for heat transfer problem in concrete structure

Abstract

In solving Partial Differential Equations (PDEs), numerical methods like Finite Element Methods (FEM) are popular choices. However, despite its popularity, FEM does pose certain limitations. For instance, it requires additional techniques when dealing with nonlinear engineering problems and inverse problems. Additionally, generating accurate predictions using FEM may necessitate a finer mesh, thereby increasing computational expenses. With increasing computing power and the emergence of deep learning, the concept of Physics Informed Neural Networks (PINNs) has gained prominence. PINNs combine the knowledge of physics and neural networks, making them a promising approach for solving PDEs. This research aims to build a PINN model to solve civil engineering-related problems, particularly focusing on heat transfer within concrete structures. It begins by constructing a PINN model for simple problems such as diffusion equations and heat transfer problem in metal rod. Subsequently, it progresses to solving the heat transfer problem in concrete slab structures. Throughout the research, it is found that PINN alone is unable to yield accurate predictions as the complexity of the problems increases. Thus, several methods or techniques such as Self-Adaptive Physics Informed Neural Networks (SA-PINN) and feature scaling will be needed to improve the model’s accuracy. Overall, the research offers valuable insights into the development of PINNs for solving PDE equations.