Development of a machine learning based automated model to predict the load-bearing capacity of circular hollow section brace members having accidental joint eccentricity

Concentrically braced members are highly effective in limiting inter-story drifts and deformations during seismic events due to their inherent lateral stiffness and load-carrying capacity. In practical seismic design, understanding and predicting the cyclic axial response of these members, where buckling and yielding states are allowed, is critical for the safety of the system. However, compression members in real structures are not perfectly straight, aligned, or concentrically loaded as is assumed in design calculations, there is always an initial imperfection. Determining the response of these members by experimental and numerical methods is very laborious and time consuming. Therefore, this study aims to design a machine learning (ML) based GUI (Graphical User Interface) to predict the load-bearing capacity of circular hollow section (CHS) brace members having accidental joint eccentricity. In this scope, 12 experimental studies and 608 finite element analyses (FEA's) were performed to train the ML models. Then, the hyperparameters were optimized by Grid Search and Random Search methods. Using the best hyperparameters, Random Forest (RF), Gradient Boosting Regressor (GBR), Multi-Layer Perceptron (MLP), Extreme Gradient Boosting Regressor (XGR), Support Vector Regression (SVR), Bagging-Boosting and Decision Tree (DT) machine learning models were trained and tested. In addition, the SHAP method was used to examine the relationship between input and output features. MLP was the most powerful prediction model with an R2 value of 0.999 in both optimization methods. All models used in the study are within reasonable error rates and the slope of the regression line is 0.970 and above in all models. According to SHAP analyses, buckling load has the highest impact and significantly affects the model output, followed by bottom eccentricity, top eccentricity and diameter traits. Buckling load has a wide distribution and has a strong positive effect on model predictions, while the effects of radius and thickness on model output are less pronounced. Finally, a GUI based on the MLP model, which is the most powerful model available to researchers working in this field, was created and the load-bearing capacity of CHS brace members were predicted.