Artificial Neural Network-Based Prediction of Compressive Strength for Mix Design Evaluation in Sustainable Expanded Polystyrene-Infused Concrete

Lightweight concrete incorporating expanded polystyrene (EPS) remains an active area of research due to its potential to produce more sustainable resource-efficient construction materials. However, identifying the optimal mix design for EPS-infused concrete typically requires extensive experimental trials, resulting in significant time, cost, and material consumption. To address this challenge, this study proposes an artificial neural network (ANN) predictive model with 5-fold cross-validation to estimate compressive strength performance and to develop mix design recommendations based on actual and predicted results. A total of 55 experimental samples were prepared and grouped into 11 batches, with the EPS volume replacement levels ranging from 0% to 50% at 5% increments. Model performance was evaluated using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient of determination (R2), and scatter index (SI), with graphical representations like predicted vs. actual plots, response plots, and residual plots, and the results were benchmarked against a multiple linear regression (MLR) model. Among the tested configurations, the 4-5-1 ANN model demonstrated the highest predictive accuracy. Furthermore, a Shapley (SHAP) analysis was conducted to interpret the model behavior and determine the relative importance of the input variables. The findings reveal that EPS content had the greatest influence on compressive strength prediction, followed by slump value, then gravel content, and finally concrete density.