Predicting The Mechanical Properties Of Engineered Cementitious Composites Using Machine Learning

Study in the field of engineered cementitious composites has developed over time within the civil engineering discipline.It has been well established that several parameters may influence the performance of the composite.Artificial neural network(ANN)algorithms are recognised as a useful tool to predict material behaviour concerning these influencing parameters accurately.The benefit and importance is the ANN models capability to predict solutions by being trained with experimental data.They essentially catalogue the performance characteristics eliminating the need to refer to tables and the requirement for additional timeconsuming testing.The predictions of the model will aid in continuing research,development and implementation of fibre composites.This project aimed to develop artificial neural network models for accurate prediction of the mechanical properties of ECC.The properties under study are the compressive strength,tensile stress at first crack,peak tensile stress,splitting tensile strain at first crack,ultimate tensile strain,flexural strength at first crack,peak flexural strength,flexural strain at first crack,and peak flexural strain.Mechanical properties such as the strain capacity are explicitly challenging to predict from only the constituents of mix designs,especially after the incorporation of multiple fibre types and combinations in the design process.Furthermore,the task of estimating outputs using artificial neural networks becomes far more challenging if the data available is relatively small.In this project,we propose and investigate several solutions to deal with this issue.In the first part of this work,data was collected from a comprehensive literature review of previous articles.Data from a total of 434 experiments of ECC was collected into the dataset.Solutions to deal with missing information in the dataset are then proposed.Implementation of each solution is carried out by fitting a linear regression model and measuring its R-square score and explained variance.Based on the results obtained,a separation and regrouping method was selected to process the collected data.The number of training instances available for each mechanical property are 227 for first crack uniaxial stress,227 for first crack uniaxial strain,284 for peak uniaxial stress,293 for peak uniaxial strain,60 for first crack flexural strength,48 for first crack flexural strain,189 for peak flexural strength,166 for peak flexural strain,and 313 for compressive strength.Next,data analysis using direct correlation study,Pearson’s R-Coefficient,and feature importance are carried out.The second part of this work highlights the development of the model.In our approach,random forests regressors trained using a sequential stacking technique with deep neural networks are combined to develop an ensemble model.This technique was inspired by previous works that dealt with small datasets.The neural network models are then developed on top of the random forest regressors to create the final ensemble model.The model shows improvement measured in M.A.E.scores over the previously developed models of Hossain et al.and L.Shi et al.who developed models specializing in PVA based ECC.The average R-square of 0.81 for all the PVA-ECC properties.The average error measured by R.M.S.E.is less than 0.0678 on all other fibre types and combinations,including hybrids.An outlier resistance test was conducted,and the results show that the combined techniques maintained a good R-square score even at different rates of missing,noised,or blemished data compared to the single output neural network solution.A software solution was built on top of the model and used to predict samples from unseen papers published in 2019 with satisfactory performance.This project has verified the ability of an artificial neural network to make accurate generalized predictions within the given domain of the supplied training data.Improvements to the generalized predictability of the neural network were realized through the selection of an optimal network configuration and training method suited to the supplied training data set.Furthermore,The developed software can be utilized to predict experimental test results of Compressive strength,Tensile stress at first crack,Peak tensile stress,Splitting tensile strain at first crack,Ultimate tensile strain,Flexural Strength at first crack,Peak flexural Strength,Flexural strain at first crack,and Peak flexural strain.This functionality is limited to the domain of the training data.It is capable of ultimately saving time and money otherwise used in conducting further testing.