Predicting the compressive strength of engineered geopolymer composites using automated machine learning

Abstract

Engineered Geopolymer Composites (EGC) offer a sustainable and high-performance alternative to traditional concrete and Engineered Cementitious Composites, boasting reduced environmental impact and enhanced mechanical properties. However, optimising EGC mix design for specific applications requires an accurate prediction of its compressive strength. This study investigates the application of Automated Machine Learning using the PyCaret library to develop reliable predictive models for EGC compressive strength. A comprehensive experimental program was conducted, testing 132 EGC specimens with varying mix design parameters, including binder ratio, silica fume content, activator ratio, water content, superplasticizer dosage, and curing method. The collected data was used to train and evaluate twenty different machine learning models. Model performance was assessed using various metrics. The top six models were shortlisted and optimised using Random Search algorithm. The models were assessed through a detailed analysis of their residual plots and learning curves. Additionally, Feature importance and SHAP analysis were employed to understand the influence of each input parameter on the predicted compressive strength. The results demonstrate the effectiveness of AML in accurately predicting EGC compressive strength, with the Gradient Boosting Regressor and CatBoost Regressor models exhibiting superior performance, achieving Mean Absolute Error (MAE) below 1.2 MPa and R² exceeding 0.96. The study highlights the potential of AML as a valuable tool for optimising EGC mix design and promoting the wider adoption of this sustainable construction material.