Predicting compressive strength of concrete with iron waste: a BPNN approach

Abstract

This study presents a comprehensive exploration into predicting the compressive strength of concrete by incorporating waste iron as a partial substitute for sand, employing a backpropagation neural network (BPNN) model. The optimized BPNN model, fine-tuned with intricate hyperparameters, demonstrates exceptional predictive accuracy, achieving an R² score of 0.9272 on the test set. Low mean squared error (MSE), Root Mean squared error (RMSE), Mean absolute error (MAE), and mean absolute percentage error (MAPE) values underscore the model's proficiency in minimizing prediction errors. The hyperparameter optimization process results in a complex neural network architecture, highlighting the intricate nature of capturing the nuances of concrete compressive strength. Visualization tools, including actual versus predicted plots and radar plots, offer clear insights into the model’s consistent excellence across various metrics. The analysis not only validates the model's precision but also provides a visually intuitive representation of its performance. Global sensitivity analysis reveals that the percentage of iron waste (‘Iron Waste (%)’) emerges as a pivotal factor, with ST and S1 values of 0.668864 and 0.643553, respectively, influencing the variability in compressive strength predictions. ‘Age of concrete’ of the concrete follows as the second most influential factor, with ST and S1 values of 0.344926 and 0.321598, respectively. This study contributes to understanding the intricate relationships between input features and concrete compressive strength, emphasizing the importance of considering the proportion of iron waste in sustainable concrete mixtures. Overall, the findings provide valuable insights for optimizing concrete formulations and advancing eco-friendly construction practices.