Prediction of buckling damage of steel equal angle structural members using hybrid machine learning techniques

This article deals with prediction of buckling damage of steel equal angle structural members using a surrogate model combining machine learning and metaheuristic optimization technique. In particular, a hybrid Artificial Intelligence (AI)-based model involving Artificial Neural Network (ANN) and Particle Swarm Optimization (PSO) was developed and calibrated for the problem at hand. For this purpose, a database concerning compression tests of steel equal angle structural members was constructed from available resources with geometry variables such as length, width, thickness, mechanical properties of materials such as yield strength and initial imperfections (i.e. residual peak stress and initial geometric imperfections) and critical buckling load of columns. The hybrid PSOANN model was adopted because its prediction capability is higher than the traditional technique - i.e. scaled conjugate gradient (SCG). Indeed, ANN trained by PSO delivered better performance in terms of RMSE, MAE, ErrorStD, R² and Slope in comparison to ANN trained by SCG, for instance. RMSE decreases from 0.141 to 0.055; MAE decreases from 0.108 to 0.042; R² increases from 0.749 to 0.959, when switching from ANN alone to hybrid PSOANN, respectively. Moreover, a Partial Dependence (PD) investigation was performed to interpret the "black-box" PSOANN model.