Prediction of flexible pavement 3-D finite element responses using Bayesian neural networks

Abstract

3-D finite element (FE) models have many advantages over traditional mechanistic tools. They allows modelling of tyre–pavement interactions, including complex viscoelastic or stress-dependent material behaviour and they capture time-dependent load interactions. However, they require significant computational resources and knowledge to run and interpret. This study introduces a surrogate model to predict 3D FE model responses using Bayesian neural networks that include model uncertainty. 850 3-D pavement FE cases were used to develop the model. Data size was increased to 1600 using data augmentation techniques. To quantify model uncertainty, Monte Carlo dropout was used as Bayesian approximation. The model was tested for sensitivity to ensure that the physical laws of the problem were not violated. In addition, the model was tested for calibration to make sure that prediction bound matched observations. Shapley Additive Explanations (SHAP) were utilised to increase the interpretability of the neural network. The highest relative error was below 15%. The model can be used to accurately predict 3-D FE pavement model responses and/or as a diagnostic tool to identify regions with sparse data that result in high model uncertainty.