Methodology for Physics-Informed Neural Networks for materials test result prediction: Case study on the Marshall test for asphalt

Physics-informed neural networks (PINNs) are increasingly used in science and engineering to facilitate predicting physical quantities in complex problems. In this paper, we present a detailed methodology for applying PINNs to predict construction material test results. We demonstrate the methodology using 2656 Marshall test results for asphalt mixtures. First, we developed the governing equations of the asphalt mixture under the Marshall test by using a viscoelastic-plastic damage model. After that, we preprocessed the dataset and performed feature engineering to propose new features, followed by computing their correlations. Next, we introduced the PINNs architecture through an inverse PINN formulation (PINNverse) while applying hard physics constraints. The resulting prediction model shows good performance, by achieving test reproducibility-adjusted R-squared values of 0.88 for Marshall Stability and 0.55 for Marshall Flow, and a prediction interval coverage of 93% for both. Furthermore, PINN formulation for the Marshall Flow performed 7.8% better than classical neural network models. To evaluate the prediction model performance, we proposed new statistical metrics that consider the reproducibility of the physical test. The proposed methodology can be applied to other material test methods and we have made it available, along with the developed web-based calculation tool through AsphaltMine.org.