Deep learning to predict the hydration and performance of fly ash-containing cementitious binders
Deep learning to predict the hydration and performance of fly ash-containing cementitious binders
复制标题
DOI：
10.1016/j.cemconres.2023.107093
复制
发表时间：
2023-03
影响因子：
11.4
通讯作者：
Taihao Han;Rohan Bhat;Sai Akshay Ponduru;A. Sarkar;Jie Huang;G. Sant;Hongyan Ma;N. Neithalath;Aditya Kumar
中科院分区：
文献类型：
--
作者：
Taihao Han;Rohan Bhat;Sai Akshay Ponduru;A. Sarkar;Jie Huang;G. Sant;Hongyan Ma;N. Neithalath;Aditya Kumar
飞灰(FA)是一种工业副产品，用于部分替代混凝土中的波特兰水泥(PC)，以减轻混凝土对环境的影响。FAs的化学组成和结构对混凝土的水化动力学和抗压强度有显著影响。由于FA的这些物理化学属性具有很大的多样性，因此开发一个通用的理论框架--因此，开发基于理论的分析模型--能够对[PC+FA]结合剂的性质产生可靠的、先验的预测一直是具有挑战性的。近年来，机器学习(ML)--它纯粹是数据驱动的，而不是源自理论基础--已经成为一种很有前途的工具，用于预测和优化包括上述粘结剂在内的复杂、异质材料的性能。也就是说，有两个问题阻碍了ML模型的广泛使用：(1)ML模型需要数千个数据记录来获取输入-输出关系，开发如此大但一致的数据库是不切实际的；(2)ML模型--尽管擅长产生预测--无法揭示材料成分/结构与其性质之间的潜在相关性。本研究采用深森林(DF)模型预测[PC+Fa]粘结剂的成分和时间相关的水化动力学和抗压强度。数据降维和分割技术--前提是从理论上理解FA的组成-结构相关性，以及PC的水化机理--被用来提高DF模型的预测性能。最后，通过对DF模型的中间和最终输出的推断，建立了一个简单的封闭形式的分析模型来预测抗压强度，并揭示了混合料设计与[PC+FFA]粘结剂抗压强度之间的关系。
Fly ash (FA) – an industrial byproduct – is used to partially substitute Portland cement (PC) in concrete to mitigate concrete's environmental impact. Chemical composition and structure of FAs significantly impact hydration kinetics and compressive strength of concrete. Due to the substantial diversity in these physicochemical attributes of FAs, it has been challenging to develop a generic theoretical framework – and, therefore, theory-based analytical models – that could produce reliable, a priori predictions of properties of [PC + FA] binders. In recent years, machine learning (ML) – which is purely data-driven, as opposed to being derived from theorical underpinnings – has emerged as a promising tool to predict and optimize properties of complex, heterogenous materials, including the aforesaid binders. That said, there are two issues that stand in the way of widespread use of ML models: (1) ML models require thousands of data-records tolearninput-output correlations and developing such a large, yet consistent database is impractical; and (2) ML models – while good at producing predictions – are unable to reveal the underlying correlation between composition/structure of material and its properties. This study employs a deep forest (DF) model to predict composition- and time-dependent hydration kinetics and compressive strength of [PC + FA] binders. Data dimensionality-reduction and segmentation techniques – premised on theoretical understanding of composition-structure correlations in FAs, and hydration mechanism of PC – are used to boost the DF model's prediction performance. And, finally, through inference of the intermediate and final outputs of the DF model, a simple, closed-form analytical model is developed to predict compressive strength, and reveal the correlations between mixture design and compressive strength of [PC + FA] binders.