Reliability Approach for Structural Safety Assessment Using Finite Element Method and Machine Learning

Abstract

Traditional methods for analysing geotechnical structures typically rely on the deterministic approach, where the primary focus is on determining factor of safety (FoS) for structure. However, this approach may yield different failure probabilities despite similar values of safety factors due to conservative assumptions and unreliable initial data. This paper integrates reliability approach, finite element method (FEM), and machine learning (ML) to enhance safety assessments in geotechnical engineering. In this paper, an example of a uniaxial compression test of specimen is considered. Numerical simulations are conducted using PAK software. The parameters of the material model are generated according to the Gaussian (normal) distribution and 10 000 calculations are created based on these distributions. Results obtained from numerical simulations are used to train a neural network, which classifies calculation as converged or diverged, i.e., a structure as stable or unstable based on the material model parameters. This approach reduces the need for an extensive number of time-consuming numerical simulations, which can be a significant resource drain in engineering applications. It can be concluded that, by coupling FEM and ML, safety of the structures can be effectively analysed using the concept of reliability.