ContactGAN development – prediction of tire-pavement contact stresses using a generative and transfer learning model

ABSTRACT

An accurate characterisation of tire-pavement contact stresses is important for pavement structural analysis and performance evaluation. The demand for rapid pavement design and analysis requires a simple, accurate, and robust approach that pavement engineers can easily use. In this study, a generative adversarial network model – ContactGAN – was developed to replace the finite element analysis (FEA), for predicting the three-dimensional (3D) non-uniform contact stress. The ContactGAN model consists of a Generator that generates synthetic stress fields, and a Discriminator that distinguishes synthetic and FEA-generated stress fields. Three finite-element (FE) generated tire contact stress datasets were used. Compared with the previously developed ContactNet model, the ContactGAN model produces a more accurate contact stress distribution in both the overall quality and the most critical values. As part of the deep learning model, a transfer learning method was developed to address the challenges of applying the model to tasks with minimal datasets. After the transfer learning approach is applied, the mean square errors (MAEs) of the 3D contact stress predictions of the cornering tire dataset and the hyper-viscoelastic tire dataset are reduced by 83.8% and 92.9%, respectively. This research delivers a data-driven approach for the rapid and accurate prediction of tire-pavement contact stresses.

KEYWORDS:

• Data-driven approach
• contact stresses
• generative adversarial network
• finite element analysis
• transfer learning

Disclosure statement

No potential conflict of interest was reported by the author(s).