Comparative Study of Machine Learning for Predicting Compressive Strength in Oyster Shell Cementitious Composites

Annual oyster production in southern Korea reaches about 300,000 tons, generating an equivalent amount of waste oyster shells. Most are illegally dumped or stockpiled along coastlines, causing serious environmental issues. This study utilized machine learning to predict the compressive strength of oyster shell cementitious composites. A total of 336 datasets were used, including 189 experimental results and 147 from published literature. Input variables were water-to-cement ratio (W/C), silica fume, blast furnace slag, superplasticizer content, and curing conditions. Algorithm selection compared the performance of Ridge Regression, Support Vector Regression, Artificial Neural Network, and Random Forest (RF), with RF exhibiting the highest predictive performance (R² = 0.8411). Ensemble algorithms including XGBoost, AdaBoost, Extra Trees, and LightGBM were optimized using GridSearchCV. Among these, LightGBM showed the best predictive capability with a mean absolute error of 3.1671, mean squared error of 17.8054, root mean square error of 4.2196, and R² of 0.9042. SHAP analysis revealed that W/C and superplasticizer were the most influential variables. Oyster shells showed a negative correlation with sand, indicating the role of oyster shells as a substitute material. Thus, cementitious composites can maintain compressive strength and serve as sustainable construction materials when waste oyster shells are incorporated with appropriate admixtures.