Optimization of bolted beam connections using surrogate models, finite element analysis, and machine learning

In this study, machine learning (ML) models were developed to optimize the bolt layout to minimize the I-beam deflection at the free end when subjected to different loads. Finite element analysis (FEA) was integrated with Latin hypercube sampling (LHS) to build comprehensive datasets. Three load cases, including vertical, horizontal and combined (vertical + horizontal), and a cantilever beam fixed with the end plate bolted to the support plate, were considered. A total of 150 data points were generated, with each data point comprising of varying bolt layouts and corresponding beam deflections. Six ML models including lasso, ridge, Gaussian process regression (GPR), support vector regression (SVR), decision tree (DT), and random forest (RF) were trained and validated. The model performance was evaluated using the root mean square error (RMSE), and coefficient of determination ( R 2 ) values, and the results were analyzed using SHapley Additive ExPlanation (SHAP) method. Subsequently, the best predicted model was used as a surrogate model for optimization, including particle swarm optimization (PSO) and genetic algorithms (GA) to obtain the best optimal bolt layout to minimize beam deflection. Finally, the optimal layout given by the surrogate model was verified against a full FEA analysis in ANSYS. The results from the model and FEA show a good match, indicating that the model accurately predict the deformation and can eliminate costly FEA simulations.