On the application of neural networks for service life prediction of steel fibre-reinforced concrete

While physics-based modelling entails the capability to describe the mechanisms and physics of a system across multiple scales, these models often are decoupled from external measurements available for specific systems of interest. On the other hand, data analysis methods allow the decoding of information on particular systems, yet such mathematical analyses struggle with context of the domain physics under study. Although it is common to have both methods positioned as antagonizing, the real challenge is to combine them, capitalizing on their strength and information. Proper integration between physics-based models and data-driven inference will enable understanding, optimising, and predicting complex physical systems around us. This study exemplifies the application of machine learning to predict corrosion in fibre-reinforced concrete. Focus is placed on current limitations concerning the application of machine learning to predict the underlying mechanisms for fibre corrosion and opportunities to integrate physics-based modelling and data-driven interference truly. Results indicate that it is possible to train neural networks with high accuracy through training examples (>99% classification accuracy), while a loss in the ability to generalize patterns and thus predict the degree of fibre corrosion is observed for samples beyond the training set (<60% classification accuracy). The absence of the ability to generalize is mainly caused by the loss of physics, such as material history dependency, physical invariance, and conservation laws, which are not naturally resolved in the current machine learning models.