Using ensemble learners and deep learning for analyzing the compressive strength of geopolymer concrete

Geopolymer is a strong candidate to replace ordinary portland cement (OPC) as a binder for construction, and geotechnical applications. The primary application of OPC is the most used construction material which is concrete. The compressive strength is a primary property of concrete. This mechanical property distinguishes concrete as it is significantly higher than other properties such as tensile strength. There are two primary constituents of a geopolymer which are a solid precursor, and an alkali activator solution. There are other parameters influencing the degree of reaction between these two constituents such as precursor composition, curing conditions, aggregate proportion, and testing period. Researchers tried to measure the correlation between the compressive strength and geopolymer synthesis parameters. The used methodologies include the conventional method of testing a parameter with the others being fixed. There were disagreements about the degree of influence of each variable that made it necessary to use multivariate analysis. Taguchi method and some machine learning models utilize multivariate systems to measure variable influence and design optimization. Disagreements were still found between studies utilizing these methods. This study analyzes data from peer-reviewed journals from different places around the world. The analyses involve preprocessing of synthesis parameters and the compressive strength of geopolymer concrete. Then, different supervised learning algorithms including, MARS, ensemble learning, and deep learning were used for building robust strength prediction models. Hyper-parameter tuning is performed for each model to improve the predictive accuracy. Ten-fold cross-validation was used to ensure bias reduction for models. The effect of the chosen hyper-parameters on the performance of the model is analyzed in terms of the quality of training, and generalization. Further evaluation was performed with sensitivity analyses to measure the influence of synthesis parameters as explanatory variables. This step is followed by training models with a selected number of synthesis parameters divided into subsets to see the reflection on the accuracy. These subsets of variables are chosen based on variable influence analysis and then used to build refined models. These refined models are built to see the effect of reducing variables on predictive accuracy with different algorithms.Applied Science, Faculty ofEngineering, School of (Okanagan)Graduat