Explainable machine learning models for probabilistic buckling stress prediction of steel shear panel dampers

Steel shear panel dampers are widely used as passive energy-dissipation devices in earthquake-resistant structures. The out-of-plane buckling stress of the core plate included in the steel shear panel damper is critical for obtaining the desired lateral force-resistant and hysteretic energy-absorbing capacities. However, the existing calculation method of buckling stress of the steel shear panel damper cannot well address the effects of the steel material’s and geometries’ inherent uncertainties. This paper intends to develop the probabilistic buckling stress prediction models of steel shear panel dampers using machine learning methods considering the steel material’s and geometries’ uncertainties. To this end, the nominal buckling stress prediction models are first developed using different machine learning algorithms based on finite element analysis results where the efficiency of the finite element models has been validated through test results. The analysis results confirm the highest accuracy of the artificial neural network (ANN) model. The SHapley Additive exPlanations (SHAP) and feature importance analysis methods are adopted for interpreting the developed prediction models. The analysis results indicate that the yielding stress of steel (fy), the height-to-width ratio (α), the width-to-thickness ratio (β), and the initial imperfection (δ) show significant influences on the nominal buckling stress of the steel panel damper while the thickness of the core plate (t) shows negligible effects. The probabilistic buckling stress prediction models are further developed based on the trained ANN model considering the combined uncertainties of steel materials and geometries by the Latin hypercube sampling method. The efficiency of the developed probabilistic buckling stress prediction models is confirmed by capturing the probability density of the numerical simulation results. The global sensitivity analysis (GSA) method is introduced for investigating the influence of the design parameters on the discreteness of the probabilistic buckling stress. It has been found from the analysis results that the uncertainties of the yielding stress of steel (fy) and the initial imperfection (δ) have significant influence, that of the height-to-width ratio (α) have limited influence, and that of the width-to-thickness ratio (β) and the thickness of the core plate (t) have negligible influence on the buckling stress of the steel panel damper. For practical application, the interactive software named “PBSSD” is developed and provided for predicting the probabilistic buckling stress by inputting the nominal design parameters of steel material and geometries.