Neural Network for Constitutive Modeling of Beam Structures

Abstract Constitutive modeling plays a vital role in simulating the stress-strain behavior of structures within various engineering applications. Traditionally, constitutive models relied on experimental tests to accurately characterize the stress-strain relationship. In recent decades, there has been a significant increase in the development of advanced constitutive laws due to the high cost associated with testing. However, numerical computations based on advanced constitutive laws heavily relies on numerical integration which is time-consuming and computationally demanding. This leads to the need to develop an alternate approach. In this study, a neural network-based constitutive model using TensorFlow was proposed to capture the non-linear relationship between stress and strain in beam structures. A case study was conducted where the neural network-based constitutive model was implemented to predict the axial force and bending moment of the pipeline subjected to ground displacement, modeled as an Euler-Bernoulli beam with large deformations. The comparative analysis with traditional numerical integration schemes was pursued to assess the performance of the proposed model. The results showed a significant reduction in computation time when dealing with large data sizes. Additionally, the impacts of the training data samples on the applicability of the neural network constitutive model were investigated. The findings of this study support that the neural network-based constitutive model serve as a more efficient tool for extensive numerical simulations in the design and structural analysis of beam structures.