Modeling Of Cement Hydration Microstructure In-Situ Development Based On Conditional Generative Adversarial Networks

As one of the most important basic building materials,cement provides important support for urban and rural infrastructure construction.In recent years,the cement industry has been transitioning towards high-end and green development,which has raised higher requirements for fundamental research related to cement.As hydration reaction largely determines the properties of cement-based materials,fundamental research on cement hydration is particularly important.However,a cement hydration reaction is an extremely complex multiscale heterogeneous process,and the complete mechanism is not yet clear,posing many challenges for related research.This article starts from the perspective of the cement microstructure which has a significant impact on hydration reaction and aims to establish an accurate pixelated model to describe the hydration development process of real cement microstructure.Taking into account the shortcomings of existing cement microstructure modeling work and the difficulty in obtaining in-situ image data of cement microstructure,this thesis mainly includes several research aspects:Firstly,to address the challenge of obtaining in-situ development time series data for cement images,this study abstracts the problem of modeling in-situ cement development under ex-situ image data as a new dynamical system modeling problem.The cycle consistency loss is extended to the continuous domain to improve mutual information between states during the hydration process,achieving in-situ development modeling.To establish a model for continuous development from discrete conditional cement image data distribution while maintaining in-situ development,this thesis proposes a continuous conditional generative adversarial network based on the auxiliary regression model.To verify the effectiveness of the model,this study conducts experiments using a dataset of real cement microstructure images.The results show that this method can accurately predict the continuous hydration process of cement with high fidelity,capture the changing patterns of different phases in cement microstructure images as the degree of hydration changes,and ensure the in-situ property of the predicted trajectory.Secondly,to address the issue of uncertainty in cement hydration prediction,this thesis introduces an information bottleneck in the generator of the conditional generative adversarial network,which enables the predicted process to retain important information and discard insignificant information.The model is built in an autoregressive form to reflect the uncertainty of hydration prediction through diverse prediction results,while ensuring that the level of uncertainty is positively correlated with the prediction time span,in accordance with the laws of physics and chemistry.Experimental results demonstrate that this method can introduce diversity and randomness in the prediction results while preserving the in-situ property of cement evolution,providing richer prediction information for cement hydration modeling.Thirdly,to address the problem of model continuity sacrificing for diversity in the autoregressive model,this thesis attempts to use stochastic differential equations(SDEs)for establishing continuous and stochastic models of the cement hydration process.To extend the use of SDEs driven by Brownian motion to the task of cement microstructure images modeling,this thesis proposes to encode the images into the information domain and use SDEs to model the states in the information domain,thus ensuring the diversity while preserving the correlation of changes between adjacent pixels.Experimental results show that although there is some information loss,this method can still meet the requirements of continuity,diversity,and in-situ modeling of the cement hydration microstructure development.