Mechanical Performance and Machine Learning Prediction of Oyster Shell Powder Concrete: A Comparative Study of GPR and Decision Tree Models
Samuel Wilberforce Offei *
School of Engineering, Accra Institute of Technology, Private Mail Bag, Accra, Ghana.
Oyindamola Ayomide Okesanjo
School of Engineering, Accra Institute of Technology, Private Mail Bag, Accra, Ghana.
Ama Annorwah Ankamah-Cofie
School of Engineering, Accra Institute of Technology, Private Mail Bag, Accra, Ghana.
Onajah Jukaudo Matthew
Department of Civil Engineering, Tamale Technical University, Tamale, Northern Region, Ghana.
*Author to whom correspondence should be addressed.
Abstract
The increasing demand for sustainable construction materials has stimulated interest in the use of waste-derived materials as partial cement replacements in concrete. Oyster shell powder (OSP), a calcium-rich waste material, offers potential for reducing cement consumption while promoting waste valorisation. However, predicting the mechanical performance of OSP-incorporated concrete remains important for evaluating suitable replacement levels and supporting material design. This study investigated the mechanical properties of concrete incorporating OSP and evaluated the predictive performance using MATLAB-based regression and machine learning techniques. Experimental compressive strength data were obtained at five OSP replacement levels (0%, 10%, 20%, 30%, and 40%) and four curing ages (7, 14, 21, and 28 days), yielding 20 average observations. Gaussian Process Regression (GPR) with a Matérn 5/2 kernel and a Fine Decision Tree model were evaluated using validation metrics. The experimental results showed that 10% OSP replacement produced the highest 28-day compressive strength of 22.12 MPa, compared with 20.24 MPa for the control mixture, representing an increase of approximately 9.3%. At 40% replacement, the 28-day compressive strength decreased to 17.06 MPa. GPR outperformed the Fine Decision Tree model, achieving a validation root mean square error (RMSE) of 0.673 MPa and a coefficient of determination (R2) of 0.92, compared with an RMSE of 1.948 MPa and a coefficient of determination (R2) of 0.32 for the Fine Decision Tree. Notably, a multiple linear regression baseline outperformed the Fine Decision Tree, indicating that complex, rule-based models overfit the limited dataset, whereas GPR's smoother inductive bias allowed for better generalisation. For splitting tensile strength, a quadratic polynomial regression was adopted because only five 28-day observations were available. The measured tensile strength peaked at 2.35 MPa at 10% OSP replacement, compared with 2.01 MPa for the control mixture, before declining to 1.67 MPa at 40% replacement. The fitted regression model achieved a coefficient of determination (R2) of 0.9355. These findings identify 10% OSP as the most promising replacement level among those investigated and demonstrate the potential of GPR for predicting compressive strength within the experimental domain. However, the limited dataset highlights the need for additional experimental observations and independent validation before the models can be generalised to broader concrete mixtures.
Keywords: Oyster shell powder, sustainable concrete, compressive strength, splitting tensile strength, gaussian process regression, fine decision tree, polynomial regression, MATLAB, machine learning