A machine learning-based study on ultimate bearing capacity of engineered bamboo beams in bending

In this study, a brief overview of the four-point bending test method for engineered bamboo beams and key experimental findings were presented. The analysis revealed that traditional approaches for predicting the ultimate bearing capacity based on stress–strain relationships in tension and compression were computationally inefficient. To address this issue, six machine learning models were developed to predict the ultimate bearing capacity of engineered bamboo beams in bending. Based on physical significance, the initially selected seven input features were optimized to four key features: width, height, span, and compressive modulus of elasticity. Among the developed models, the model based on artificial neural networks (ANN) achieved the highest accuracy, with a coefficient of determination of 0.98, a root mean square error of 2.56 kN, and a mean absolute error of 1.79 kN. The ANN model was followed by extreme gradient boosting, k-nearest neighbors, support vector regression, and decision tree models, whereas the Bayesian linear regression model showed the lowest performance. Input feature importance analysis based on the Shapley additive explanations approach revealed that geometric dimensions had a more significant impact on the ultimate bearing capacity than the intrinsic mechanical properties of the material. A physically interpretable theoretical model was developed by integrating trained ANN outputs with a fitting approach, offering a more efficient, practical, and scalable solution for evaluating the ultimate bearing capacity of engineered bamboo beams in bending.