An ensemble method for predicting the mechanical properties of strain hardening cementitious composites. (7th June 2021)

• Record Type: Journal Article
• Authors: Altayeb, Mohamedelmujtaba; Wang, Xin; Musa, Taha Hussein
• Is Part Of: Construction & building materials. Volume 286(2021)
• Publication Date: 2021-06-07
• Subjects: Artificial neural networks -- Ensemble methods -- Strain-hardening cementitious composites; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2021.122807
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: Reliably predicting the properties of SHCC using models build from small datasets with partially missing data. An FDNN model that predicts the mechanical properties of SHCC regardless of the type of fibre reinforcement. The FDNN model outperforms current neural network models in predicting SHCC's mechanical properties. A potentially general model for studying SHCC behaviour. Abstract: In recent years, Artificial Neural Networks (ANN) models have proven effective in learning to predict material properties only from data. In the study of strain-hardening cementitious composites (SHCC), conducting laboratory experimentation to collect this data is expensive, and to reach reliable accuracy, tens of thousands of instances may be required. In this paper, a forest deep neural network (FDNN) ensemble model is presented. The proposed model combines random forest regressors with artificial neural networks to predict the cementitious composite's mechanical properties. The FDNN model can provide reliable predictions after training on relatively small datasets of different classes of SHCC. The FDNN outperforms previously developed artificial neural networks model architectures build specifically for polyvinyl alcohol (PVA) based SHCC. The model achieves an average root means square error of 0.08, and its predictions for several specimens are provided and discussed in this work.