Machine Learning–Based Surrogate Optimization of Buckling Delayed Shear Link Dissipators
Please login to view abstract download link
The numerical modelling of metallic seismic energy dissipation devices involves strongly nonlinear behavior, such as large displacements and deformations, cyclic plasticity, and complex contact conditions. Buckling Delayed Shear Link (BDSL) dissipators are a representative example, where small geometric variations can significantly affect stiffness, energy dissipation and damage mechanisms. Although high-fidelity Finite Element Method (FEM) models can accurately reproduce their response [1], their computational cost limits their use in large-scale optimization problems. This work presents a surrogate-assisted numerical optimization aimed at identifying optimal designs of BDSL dissipators. A three-dimensional FEM model, calibrated against experimental cyclic tests, is used to generate an initial parametric dataset by varying key geometric parameters. On this basis, Machine Learning (ML) surrogate models and Radial Basis Function (RBF) approximations are trained and compared to predict deformation and damage-related response variables. The surrogate models are embedded into an evolutionary optimization strategy based on Differential Evolution [2], in which the geometric configuration is obtained by minimizing an objective function that combines shear windows distortion and damage-related indicators. The optimization process is formulated as an iterative scheme, where selected geometries are re-evaluated using FEM and subsequently incorporated into the training dataset, enabling adaptive refinement of both ML and RBF surrogate models. The proposed optimization strategy accurately reproduces the FEM response of BDSL devices while drastically reducing the computational cost through a significant reduction in the number of FEM evaluations. The results demonstrate the effectiveness of surrogate modeling combined with evolutionary optimization for the numerical design of complex structural systems with strongly nonlinear behavior.