Parametric analysis and optimized shear capacity prediction of stiffened steel plate girders using hybrid machine learning algorithms

Abstract

Structural engineers face a hard time accurately predicting the level of shear resistance reinforced steel plate girders. The reason for this is the complicated network of nonlinear interactions that exists between the qualities of the material, the shape of the girder, and the behavior of the post-buckling tension field. Basler's and Cardiff's tension field theory is the basis for much of the empirical formulas used in long-standing design standards like IS 800:2007, AISC 360–16, and Eurocode3. People often don't think about how flexible and strong the flanges are after they bend. This paper presents a Hybrid ML framework that can predict shear resistance close to accurate and lets for in-depth parametric analysis. A dataset from 314 experiments and using numerical simulations. Thirteen non-dimensional measures were included to show how these structures work in conjunction careful feature engineering. These metrics include web slenderness (dw/tw), flange stiffness (bf/2tf), and moment of inertia ratios of flanges and web. The study looked at nine different regression methods, including Random Forest, XGBoost, CatBoost, AdaBoost, Decision Tree, Multilayer Perceptron, Artificial Neural Networks, LightGBM, and Gradient Boosting Machines. Eight different ways were tried by changing the models' hyperparameters to see how well they could predict. Besides Grid Search, Successive Halving, and Random Search there were other advanced bio-inspired metaheuristics such as Genetic Algorithm, Particle Swarm Optimization, Artificial Gorilla Troops Optimizer, Ant Colony Optimization, and Colliding Bodies Optimization were used to find the best solution. A new Stacking Ensemble Regressor was developed using a Linear Regression meta-learner. It incorporates the best optimized basic learners, like GA-tuned RF and AGTO-tuned CatBoost. This ensemble has a Root Mean Squared Error (RMSE) of 24.65 kN and a coefficient of determination (R^2) of about 0.994. The model was much ahead of the standalone models and the analytical models made by Cardiff and Hoglund. SHAP (SHapley Additive exPlanations) and PDP were used in a study on interpretability to confirm the results of theoretical mechanics. The most important factors recognised were web slenderness and web thickness. Lastly, a Python Graphical User Interface (GUI) was developed to connect complicated computational models to real-world design uses.