Machine Learning for collapse structural analysis of Hollow Elements

The prediction of the collapse of hollow structural elements is traditionally based on nonlinear finite element (FE) analyses, which are accurate but computationally demanding, making them unsuitable for extensive parametric studies or rapid assessments. This thesis develops an automated framework that reduces the number of required numerical simulations with surrogate models based on machine learning. The workflow integrates Sobol sampling for the Design of Experiments, high-fidelity Finite Element Model (FEM) simulations in Abaqus, automatic extraction of collapse indicators (Area Reduction, Diameter Reduction) and vectorial responses (Area Profile), and finally the construction of surrogate models using Gaussian Process Regression (GPR) and Multi-Layer Perceptrons (MLP). The results indicate that the surrogates reproduce the reference simulations with engineering-level accuracy, reducing computation times from 80–700 seconds per run to approximately 0.1 seconds per prediction, achieving a speed-up of three to four orders of magnitude. GPR additionally provides uncertainty estimates, useful for engineering applications, while MLP proves more suitable for larger and more complex datasets. Although the work is based on simulated data and has room for improvement in terms of numerical stabilization and uncertainty calibration, it provides a validated, extensible, and reproducible tool. This creates opportunities for future developments such as sequential sampling, hybrid model approaches, with potential impact on design and predictive maintenance.