Neural network predictions of the simulated rheological response of cement paste in the FlowCyl

For the past decades, computational fluid dynamics (CFD) simulations have been shown as a promising approach for understanding the complex flow behavior of concrete. However, their application is often limited due to the computationally heavy analysis. In this study, two artificial neural networks, multi-layer perceptron and radial basis function, are trained by results of a CFD model that simulates the cement flow in the FlowCyl equipment. Both models were investigated for predicting single values of volume loss over a predetermined duration as well as the full volume loss versus time curves. The results show that after training the neural networks can accurately substitute the predictions of the CFD model for both single values and the full curves. For the multi-layer perceptron, accurate predicts are even obtained after substantial reducing the training data, which illustrates that a coupling between a CFD model and a neural network can significantly decrease the overall calculation time.

Automated summary

This study utilizes artificial neural networks to substitute the complex analysis of cement flow conducted by a CFD model, and the results show accurate predictions for both single values and full curves after training, with a significant decrease in calculation time achieved through coupling.