真实图像引导的水泥水化微结构三维演化反向智能建模与性能评估

Computer modeling and simulation for cement hydration and microstructure evolution have been proven to be a useful tool for understanding the chemical mechanism of cement hydration, predicting the properties of hydration process, researching the relationship between microstructure and performance, and improving the design of high performance cement material. However, due to the complexity of cement hydration mechanism and paste microstructure, there’s no fix-positioned hydration model and performance estimation method, which close to real development of microstructure from both structure and process. Scientists failed to establish the relatively real initial status for virtual microstructure, failed to define a development method that reflects the real hydration process of microstructure, and also failed to establish an accurate and logical estimation method for indicators of virtual microstructure. Guided by real three dimensional and four dimensional microstructure images, this project will reversely build three dimensional fixed microcsructure development and performance estimation model, which closes to real structure and process, using intelligent method. The project first extracts Markov Random Fields model that can be customized and describes the initial microstructure using evolutionary computation. Then, based on obtained four dimensional data, under the environment of GPGPU clusters, the project builds three dimensional cellular automata, whose cell’s behavior is controlled by deep neural networks, to simulate development of microstructure. The compressive strength and porosity are estimated by image features of virtual microstructure directly.

水泥水化过程及其微观结构演化的建模与模拟是揭示水泥水化原理、预测水化过程、研究微结构与性能之间的关系及改进高性能水泥材料设计的有效手段。然而，由于水泥水化机理和泥浆微观结构的极端复杂性，目前尚不存在从结构和过程两个方面都接近真实情况的微结构定点演化模型和性能评估方法。人们未能为虚拟微元建立较为真实的初始三维状态、未能定义反映微观结构真实演化过程的演化方法、也没能为虚拟微观结构建立一套准确合理的指标计算方法。本课题力求以真实三维与四维微观结构图像为导向，通过智能算法反向建立接近真实结构与过程的微元三维定点演化与性能评估模型。课题首先利用进化计算萃取可以自由定制并描述初始微观结构的马尔可夫随机场模型，然后在先期研究已获得四维数据的基础上在GPGPU机群环境下生成以深度神经网络为细胞行为核心的三维细胞自动机来模拟微结构的演化，并基于虚拟微元的图像特征直接估计抗压强度和孔隙率等宏观指标。

水泥及其基材料在日常生产活动中占据着极其重要的地位。水泥水化过程及其微观结构演化的建模与模拟是揭示水泥水化原理、预测水化过程、研究微结构与性能之间的关系及改进高性能水泥材料设计的有效手段。本课题力求以真实三维与四维微观结构图像为导向，通过智能算法反向建立接近真实结构与过程的微元三维定点演化与性能评估模型。在基金项目的支持下，课题组在 《IEEE Transactions on Neural Networks and Learning Systems》， 《IEEE Transactions on Systems, Man, and Cybernetics: Systems》，IEEE SMC等刊物与会议上发表了相关学术论文57篇。本项目基于真实三维数据，反演了面向水泥水化的微观结构生成模型；构建了人工神经网络来仿真微观结构演化；基于微观图像实现了水泥宏观指标的直接评估。针对课题中所涉及到的反向演化、函数生成、神经网络、参数优化、结构分析等问题，进行了深入的理论和算法研究。此外，在神经网络、参数优化等理论研究的结果也被推广到其他典型问题。在水泥水化过程建模与仿真方面形成了以反向建模与真实图像为核心的自身的研究特色。