A comparative study of NGBoost and traditional machine learning models for prediction of compressive strength of geopolymer concrete

While several studies have previously explored the prediction of compressive strength in geopolymer concrete, many suffer from limitations in feature selection, model generalizability, and prediction accuracy. This invention aims to enhance the prediction process by employing advanced machine learning algorithms capable of capturing complex, non-linear relationships between mix design parameters and compressive strength outcomes. To realize this objective, a dataset consisting of 276 geopolymer concrete mixes and their corresponding 28-day compressive strength values was compiled. Input features were selected based on two key criteria: their proven relevance in prior literature and their statistical significance in model performance. Multiple regression models—including Linear Regression, Decision Tree, Random Forest, Gradient Boosting, XGBoost, and NGBoost—were implemented and evaluated. Through trial-and-error, optimal hyperparameters such as the number of training epochs and k-fold values for cross-validation were determined. Model performance was assessed using standard evaluation metrics (R, RMSE, MAE, MSE), and further validated via score-based analysis. The model’s adaptability was tested using an independent secondary dataset. The results confirm that the NGBoost model achieved the most accurate predictions among all tested models, outperforming traditional approaches in both accuracy and consistency. This invention offers a scalable and reliable solution for predicting compressive strength, significantly reducing the need for physical trial mixes and enabling efficient, data-driven mix design in geopolymer concrete applications.