Interpretable Machine Learning For Concrete Compressive Strength Prediction: A Neural Network Model With SHAP-Based Explainability And Robust Multi-Metric Validation

Abstract

Accurate prediction of concrete compressive strength is crucial for efficient structural design, sustainable material optimization, and reliable quality control. While data-driven models, particularly artificial neural networks (ANNs), have shown superior predictive capability over traditional empirical methods, their widespread adoption in engineering practice is often hindered by their "black-box" nature and a lack of comprehensive, interpretable validation. This study develops a robust, fully connected neural network model to predict the 28-day compressive strength of concrete from eight key mix design parameters. The model, trained on a dataset of 1,133 mixtures, demonstrates excellent performance, achieving a test set coefficient of determination (R²) of 0.865, a root mean squared error (RMSE) of 5.91 MPa, and a mean absolute error (MAE) of 4.49 MPa. Crucially, the work transcends standard predictive analytics by integrating advanced model-agnostic interpretability techniques. Shapley Additive exPlanations (SHAP) and permutation importance analyses are employed to quantify and visualize feature contributions, revealing that the model's logic aligns with established concrete science: cement content and curing age are identified as the dominant positive factors, while water content exhibits a strong negative influence. The role of supplementary cementitious materials (slag and fly ash) is shown to be complex and context-dependent. A suite of six statistical metrics (MSE, RMSE, MAE, MAPE, R², CVRMSE) and detailed error distribution analysis provide a transparent, multi-faceted assessment of model accuracy and generalization. The results confirm that the proposed ANN is not only a high-fidelity predictive tool but also an interpretable model whose learned relationships validate domain knowledge, thereby bridging the gap between computational power and engineering trust for advanced concrete mix design.