Performance Comparison of Artificial Neural Network and Random Forest Models for Predicting the Compressive Strength of Fibre-Reinforced GGBS-Based Geopolymer Concrete Composites

Fibre reinforcement is essential in the geopolymer concrete (GPC) matrix to enhance deformation resistance and mitigate fracture propagation under tensile and bending stresses. However, a significant challenge is encountered in accurately forecasting the compressive strength of GPC reinforced with fibres. Recently, the application of machine learning (ML) models has proven effective in accurately forecasting GPC characteristics. However, due to its complexity and the scarcity of data available to researchers, developing the strength prediction technique for fibre-reinforced GPC remains in its infancy relative to traditional concrete. Despite the extensive research on estimating compressive strength in GPCs, limited studies are available on fibre-reinforced geopolymer concrete (FRGPC) due to the complex underlying phenomena involved. Therefore, this research evaluates the effectiveness of random forest (RF) and artificial neural network (ANN) algorithms to predict the compressive strength of FRGPC produced using ground granulated blast-furnace slag (GGBS). The study encompasses a detailed dataset of 110 data points with 16 input parameters comprising various mix proportions of GGBS-based GPC and curing conditions with varying types and percentages of reinforcing fibres. The performance of both ANN and RF models is assessed using a range of performance indicators, including MAE, MSE, RMSE, MAPE, and R², to evaluate their accuracy and generalization capabilities. K-fold cross-validation is applied to both models to prevent overfitting and to optimize their performance for the best output. The results demonstrate that both ANN and RF models exhibit promising performance in predicting the compressive strength properties. These techniques demonstrated strong predictive capabilities across various evaluation metrics. However, the RF algorithm revealed more accurate predictions of compressive strength than the ANN model. The models showed remarkable R² values of 0.906 and 0.902 and RMSE values of 3.822 and 5.232 for the RF and ANN models, respectively. The specimen’s age, the ratio of alkaline solution to GGBS, and the sodium hydroxide dosage emerged as critical factors significantly influencing compressive strength, as indicated by the sensitivity analysis results.