Research On Computational Intelligence Method And Key Technologies Of Cement Hydration Modeling

There are many physical changes and chemical reactions in cement paste which mixed with water and sands during the process of hydration and hardening. The internal complexity and time-varying characteristics lead to some challenging problems for cement materials. Therefore, new scientific methods and efficient tools are needed to improve the research of high-performance cement material.With the rapid development of computer and computing technology, a new computational materials science which is the interdisciplinary science of computer science and material science has borned. It is increasingly becoming an important branch of materials science and computer science. computer prediction of material properties, computer simulation of the evolution of materials, computer modeling of materials design and technology have been the main research objective of computational materials science. According to computer modeling, the best combination of materials components, structure and parameters can be obtained by research the materials structure, consitute and micro change in physical and chemical process. That is to say, its object is the development of advanced materials. It combines advanced computer modeling technology with traditional experimental approaches, established the computational model for cement hydration, provided new scientific methods and techniques approach. It is of profound theoretical and practical significance for the development of high-performance cement materials and improvement of product quality.Although a lot of progress has been made, building models manually has a high degree of difficulty due to the extreme complexity of the process of cement hydration. This Dissertation established the the combination of computational intelligence and traditional experimental approach from three aspectes:kinetics modeling, strength prediction and 3D microstructure modeling for hydration process.(1) Automatic Derivation of Early-age hydration kinetics for Portland cementThe early-age hydration of Portland cement paste has an important impact on the formation of microstructure and development of strength. However, manual derivation of hydration kinetic equation is very difficult because there are multi-phased, multi-sized and interrelated complex chemical and physical reactions during cement hydration.In this research, the kinetics model was established automatically from early-age hydration of time-series data for Portland cement. Two automatic modeling methods are proposed. The early-age hydration kinetic models are reversely extracted automatically from the observed time series of hydration degree of Portland cement using gene expression programming and flexible neural tree. Gene expression programming algorithm are used in both of these two methods for the evolution of expression form and flexible neural tree structure, respectively. During interation, for each expression form and flexible neural tree structure, the particle swarm optimization method are used to optimize adjusting parameters. These approach intereated until the best kinetic model and corresponding adjusting parameters are found. Furthermore, in order to reduce the computing time, GPUs are used for acceleration in parallel. Studies have shown that according to the generated kinetic model, simulation curves of both of these two approaches for early-age hydration are in good accordance with the observed experimental data. Furthermore, they still have good generalization ability even changing chemical composition, particle size and curing conditions. In compared to each other, the FNT kinetic model has more approximation accuracy. However, the extracted kinetics equation needs less number of parameters.(2) Concrete grade detection using floating centroid mehod based neural network classifierConcrete is the most important cement-based materials. The hydration of cement contained in concrete will affect its development of mechanical properties directly. However, traditional concrete grade detection method will consume many materials and time.This paper classified concrete grade using artifical neural network and proposed a novel classification method-floating centroid method to improve the accuracy of neural network classifier. Floating centroid method removes the fixed centroid constraint and increases the chance to find optimal neural network. Experimental results illustrate that the proposed method has favorable performance for concrete grade classification especially with respect to the training accuracy, generalization accuracy and average F-measures.(3) 3D microstructure modeling for cement paste hydration processFinally, the paper presents a real three-dimensional microstructures oriented modeling method for cement paste hydration process. This method consists of three parts:data acquisition of cement hydration, cellular automata machine model for three-dimensional evolution and neural network model for strength calculation. The data acquisition stage of cement hydration provided training data set for three dimensional evolution and neural network model for strength prediction. Three dimensional evolution iterate according to ceullar automation from an iniational image. Then, the compression strength in each step of iteration can be obtained directly using neural network model for strength predictionThe micrometer level three dimensional microstructure are obtained using micro-CT. At the same time, its corresponding compression strength are measured. The training data is obtained by image enhancement, gray level calibration and three dimension registration. Then, this data is used to guide the generation of cellular automation. model and neural network strength prediction model. When the models have been determined, given an initial three dimension image, the three dimensional microstructure can be simulated in each day by the interation of cellular automation. Then, extract the description features of the simulation results. The strength compression of simulated microstructure can be calculated from neural network prediction model.