Enhancements to a Deep Learning-Based Cellular Automaton for Cement Microstructure Evolution Modeling

Cement is a crucial construction material, and its physical properties are determined by the evolution process of its microstructure. It would be of great significance to establish an evolving model of cement hydration to obtain its physical properties at different ages. This would aid in the study of high-performance cement and provide production guidance. However, the understanding of cement hydration mechanisms is not yet comprehensive, which poses limitations on modeling the hydration process of cement. Considering that cement hydration is a localized reaction that expands to the overall system, the concept of cellular automaton is well-suited for modeling such a process. Machine learning, on the other hand, can learn the evolution rules of cellular automaton based on existing data on cement evolution. Therefore, the model based on cellular automaton combined with machine learning to simulate the hydration process of cement was proposed, but he model is not very effective and can only capture information within small neighborhoods. To improve the model, we have selected an appropriate receptive field size that allows the extraction of more suitable features without violating the local evolution rules of cement. This enables more effective learning of the evolution rules of cement. Furthermore, to address the issue of data perturbation in the original dataset and the model can only use small neighborhoods, we have applied batch normalization to normalize the input data. By incorporating these two methods, the performance of model has been improved.