Comparative Evaluation of Traditional and Deep Learning Based Machine Learning Models for Concrete Compressive Strength Prediction

Atif Khan, Afsar Ali, Syed Saqib Mehboob — Thu, 08 Jan 2026 00:00:00 +0800

Machine learning is widely used to predict concrete compressive strength because it captures nonlinear interactions among binders, aggregates, water content, and chemical admixtures. Tree-based ensemble models such as Random Forest and XGBoost often achieve high numerical accuracy; however, their discrete decision splitting mechanisms inherently produce stepwise response trends that may not reflect the smooth and continuous nature of material behavior. To overcome this limitation, this study proposes a hybrid CNN1D–BiLSTM model that integrates one dimensional convolutional neural networks for localized feature extraction with bidirectional long short term memory units to learn long range compositional dependencies and gradual transitions in mixture design. Linear Regression, Random Forest, XGBoost, and the proposed model were evaluated using 1,030 concrete mixtures from the Yeh dataset. Model performance was assessed using the coefficient of determination (R2), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Nash Sutcliffe Efficiency (NSE), Mean Bias Error (MBE), and Percent Deviation Degree (PDD) to provide a comprehensive evaluation of predictive accuracy and systematic bias. The CNN1D–BiLSTM achieved the best testing performance with R2 and NSE of 0.9324 and RMSE of 4.17 MPa, while exhibiting minimal systematic bias as indicated by low MBE and PDD values. In addition to strong numerical performance, the proposed model generated smooth partial dependence trends consistent with hydration kinetics and continuous strength development. Shapley Additive Explanations (SHAP) analysis confirmed stable and physically meaningful contributions of key mixture components, demonstrating that the hybrid framework maintains high predictive capability while preserving realistic response behavior across varying mixture compositions.

Journal of Architectural Environment & Structural Engineering Research

Comparative Evaluation of Traditional and Deep Learning Based Machine Learning Models for Concrete Compressive Strength Prediction