Systematic Analysis of the Impact of Data Preprocessing Techniques on Machine-Learning Model Performance: A Case Study of a Compressive Strength Prediction Model for Geopolymer Concrete

Abstract

Geopolymer concrete (GPC) is emerging as a sustainable alternative to ordinary Portland cement (OPC) concrete. However, developing mix designs for GPC presents unique challenges due to the variability in fly ash properties and the selection of appropriate alkaline activators. Traditional experimental and statistical methods often fall short in predicting the compressive strength of GPC. Recently, machine learning (ML) models have gained popularity in this area. This study evaluates the impact of various data preprocessing (DP) techniques namely missing data imputation, outlier detection, and normalization of the performance of an artificial neural network (ANN) model for predicting the compressive strength of GPC. The study also assessed the influence of feature selection using an ANN model considering common input variables identified in the literature. Results indicate that the chemical oxide content of fly ash, NaOH solid content, and total water content are critical to accurate strength prediction. Furthermore, curing temperature and time significantly enhances model accuracy. The study found that removing missing data and using median outlier detection improved model performance, while proper hyperparameter tuning markedly enhanced prediction accuracy. Validation through laboratory experiments confirmed the model’s reliability, demonstrating its potential for advancing GPC mix design.