Predicting Concrete Compressive Strength: A Comparative Analysis Of Artificial Neural Networks And Adaboost For Enhanced Generalization Performance

Abstract

Introduction: The accurate prediction of concrete compressive strength is critical for structural design and efficiency. Traditional testing methods are time-consuming, creating a demand for reliable machine learning (ML) models. This study compares the predictive performance and generalization capabilities of an Artificial Neural Network (ANN) and an AdaBoost algorithm for concrete strength forecasting, incorporating SHAP analysis for enhanced model interpretability. 

Methods: Using a dataset of 1030 concrete mixtures, models were developed and hyperparameter-tuned. The ANN was configured with a single hidden layer (100 neurons, tanh activation), while AdaBoost used 1000 estimators. The dataset was split 80-20 for training and testing, with performance evaluated using R², RMSE, MAE, and MAPE. K-fold cross-validation and SHAP analysis were conducted to assess model stability and feature interpretability. 

Results: Both models achieved a test R² of 0.84. However, AdaBoost exhibited significant overfitting, indicated by a near-perfect training R² (≈1.0) and a higher test MAPE (22.86%) compared to the ANN's consistent R² (0.84 on both sets) and lower test MAPE (17.17%). SHAP analysis revealed fundamentally different feature importance patterns: AdaBoost showed disproportionate reliance on Blast Furnace Slag with wide value dispersion indicating instability, while ANN demonstrated balanced, physically consistent relationships with cement and age as primary predictors. 

Discussion: The ANN model demonstrated superior generalization and robustness by effectively learning underlying data patterns without memorization, making it more reliable for practical applications than the overfitted AdaBoost model. SHAP analysis provided crucial insights into model decision-making processes, validating ANN's alignment with concrete science principles while revealing AdaBoost's sensitivity to specific dataset characteristics.