Enhancing prediction accuracy of portland cement compressive strength at low cost using machine learning and chemometric techniques

Abstract

Portland cement is essential in civil construction, serving as the basis for concrete in modern infrastructures. The 28-day compressive strength is a critical factor in ensuring the durability of structures and is influenced by factors such as composition, proportions, curing process, and environmental conditions. Accurately predicting this strength is essential to optimize produc- tion and ensure concrete quality. Traditional prediction methods, such as MLR and PLS, are limited to complex systems with many variables. In this context, artificial intelligence (AI) techniques, such as artificial neural networks (ANNs), emerge as promising alternatives, capa- ble of modeling non-linear relationships between cement components and their strength. This study developed an ANN model, using a Multilayer Perceptron (MLP), to predict the 28-day compressive strength of Portland cement. The research sought a time- and cost-efficient pre- diction method to capture the complex interactions between cement components and curing conditions. The process included data collection and preprocessing, network architecture selec- tion, hyperparameter selection, model training, and validation. The results showed that the MLP achieved coefficient of determination (R2) of 0.9851, mean square error (MSE) of 0.8744, root- mean-square error (RMSE) of 0.9351, mean absolute error (MAE) of 0.7360 and mean absolute percentage error (MAPE) of 0.7360%, outperforming traditional methods. This demonstrates the potential of ANNs as a robust tool for predicting cement strength, optimizing production processes, and quality control. Validation was performed with independent data, confirming the superiority of the model.