The complexity of the three-dimensional operation and the testing limitations of reinforced concrete (RC) members under torsion have long been difficult to explore torsion mechanisms and predict the performance of torsion. Although existing design code based on the Space Truss model provides calculation methods, the stability and accuracy of practical application is not sufficient, and estimating torsional strength remains a challenging task. On the other hand, the application of machine learning offers new possibilities to solve this problem, as it can automatically learn the inherent relationship between RC members and torsional capabilities without relying on complex equilibrium equations.
Therefore, Shenggang Chen and other researchers from China University of Mining and Technology, China Nuclear Power Design Co. Ltd., Beihang University, and China Construction First Group Corporation Limited, collaborated on a study entitled “Predicting the twisting capacity of reinforced concrete members through data-driven machine learning models.”
In this study, we first constructed a comprehensive database containing 287 RC members who failed under twisting through an extensive literature review, covering both solid and hollow specimens with a wide range of parameter values. Next, four machine learning algorithms – back-propagating artificial neural networks (BPANN), support vector machines (SVM), random forest regression (RFR), and extreme gradient boost machines (XGBM) – were trained and tested using a 10x cross-validation method. The predictive performance of these machine learning models was not only compared to each other, but also compared to the calculation results for three mainstream design codes: GB 50010, ACI 318-19, and Eurocode 2. Furthermore, sensitivity analysis was performed on the optimal model to investigate the effect of input variables on the trinia capacity of RC members.
The findings showed that machine learning models generally achieve better predictive performance than traditional design code. Of the four machine learning models, the XGBM model showed the most favorable predictive effect at r^{2} = 0.999, RMSE = 1.386, MAE = 0.86, and \bar{\lambda} = 0.976, followed by the RFR model. In contrast, the prediction accuracy of the SVM model was lowest. For design codes, the GB 50010 slightly overestimated the torsional capacity, while the ACI 318-19 and Eurocode 2 underestimated it. Especially ACI 318-19. Sensitivity analysis revealed that concrete strength was the most sensitive input parameter affecting the reliability of the predictive model, followed by lateral and long reinforcement ratios, total reinforcement ratios, section width, and stirring intensity, but section height had little effect. These findings provide a general explanation of the application of machine learning methods in solving torsion problems and provide new tools for exploring the torsion mechanisms of RC members.
“Predicting the twisting ability of reinforced concrete members through data-driven machine learning models” has been written by Shenggang Chen, Congcong Chen, Shengyuan Li, Junying Guo, Quanquan Guo, and Chaolai Li. It is open to the front. struct. Siving. Eng. 2024, 18(3):444–460, and the full text can be accessed at https://doi.org/10.1007/S11709-024-1050-x.
