Machine-learning-based design of high strength steel bolted connections

Abstract

For the design of high strength steel bolted connections, all existing standards adopt the same framework – (i) calculating the design resistance for each potential failure mode and (ii) defining the final design failure load as the minimum of the design resistances calculated from all the potential failure modes. However, this framework has been found to be tedious and also incapable of considering the complex nature of connection behaviour, in particular the transition of different failure modes (e.g., net section fracture, bearing and block tearing), and thus leads to inaccurate failure load and mode predictions. To address the aforementioned shortcomings, this paper presents a more accurate and reliable predictive framework based on machine learning. Firstly, a database including 543 experimental and numerical data was collected. Then, regression models for failure load predictions and classification models for failure mode predictions were developed based on eight machine learning algorithms, including Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbour, Adaptive Boosting, Light Gradient Boosting Machine, Extreme Gradient Boosting and Cat Boosting. The performance of the developed models was assessed based on the collected data, indicating that the Support Vector Machine models led to the best predictions of failure loads and modes. The Support Vector Machine models were then compared with existing design standards and shown to yield substantially improved failure load and mode predictions for high strength steel bolted connections. Specifically, the mean ratio of experimental and numerical to predicted failure loads from the machine-learning-based approach is equal to 1.00, while the corresponding mean ratios from the design standards range from 1.10 to 1.39. The machine-learning-based approach is capable of accurately predicting 97.2% of the total failure modes, while the design standards can only accurately predict 67.9%–85.3% of the total failure modes.