Micromechanical and time dependent properties of cement paste can be predicted based on the microstructure by using analytical or numerical models. Herein, we propose an alternative approach for predicting the creep modulus of cement paste based on deep convolutional neural…
Micromechanical and time dependent properties of cement paste can be predicted based on the microstructure by using analytical or numerical models. Herein, we propose an alternative approach for predicting the creep modulus of cement paste based on deep convolutional neural network (DCNN). The DCNN is trained using numerical simulation data obtained by the microscale lattice model, resulting in a database with more than 18000 samples. Then, 3 different DCNN architectures are built to learn from (part of) the database. Finally, the accuracy of DCNN prediction is tested on unseen samples. The proposed DCNN architectures can achieve excellent 2 accuracy on the testing set, with the R higher than 0.95. Furthermore, the distribution of creep moduli predicted by the DCNNs coincides with the original dataset. Further analyses of the feature maps show that the DCNNs can correctly capture the local importance of different microstructural phases on the predicted creep moduli. Therefore, it was concluded that a well-trained DCNN allows prediction of creep moduli based on microstructural images as input, which is computationally much more efficient compared to image segmentation and numerical simulation methods commonly used today. Of course, computational demands for training the network may be significant, but are needed in principle only once.