Suitability of Mechanics-Based and Optimized Machine Learning-Based Models in the Shear Strength Prediction of Slender Beams Without Stirrups

Accurate shear capacity estimation for reinforced concrete (RC) beams without stirrups is essential for reliable structural design. Traditional code-based methods, primarily empirical, exhibit variability in predicting shear strength for these beams. This paper assesses the effectiveness of mechanics-based and optimized machine learning (ML) models for predicting shear strength in stirrup-less, slender beams using a dataset of 784 tests. Seven ML models—artificial neural network (ANN), support vector machine (SVM), decision tree (DT), random forest (RF), AdaBoost, gradient boosting (GBR), and extreme gradient boosting (XGB)—were compared against three mechanics-based models: the Tran’s NLT Model (2020), the Multi-Action Shear Model (MASM), and the Compression Chord Capacity Model (CCC). Among the ML models, XGB and GBR demonstrated the highest predictive accuracy, with coefficients of determination (R2) of 0.974 and 0.966, respectively, indicating strong correlation with experimental data. Performance metrics such as mean absolute error (MAE) and root mean squared error (RMSE) showed that XGB and GBR consistently outperformed other models, yielding lower error margins. Statistical analysis revealed minimal bias and variability in the predictions of XGB and GBR, with a coefficient of variation (CoV) of 14%, ensuring high reliability. The NLT model, the most accurate of the mechanical-based models, achieved a mean of 1.02 and a CoV of 16% for its model error, demonstrating reasonable prediction reliability but falling behind XGB and GBR in accuracy. With Shapley additive explanations (SHAPs), the beam width and depth were identified as primary predictors of shear strength, providing critical insights for enhancing design and construction practises.